mirror of
https://github.com/qurator-spk/eynollah.git
synced 2026-06-18 19:09:15 +02:00
Compare commits
No commits in common. "main" and "v0.0.8" have entirely different histories.
120 changed files with 6750 additions and 24654 deletions
28
.circleci/config.yml
Normal file
28
.circleci/config.yml
Normal file
|
|
@ -0,0 +1,28 @@
|
|||
version: 2
|
||||
|
||||
jobs:
|
||||
|
||||
build-python36:
|
||||
docker:
|
||||
- image: python:3.6
|
||||
steps:
|
||||
- checkout
|
||||
- restore_cache:
|
||||
keys:
|
||||
- model-cache
|
||||
- run: make models
|
||||
- save_cache:
|
||||
key: model-cache
|
||||
paths:
|
||||
models_eynollah.tar.gz
|
||||
models_eynollah
|
||||
- run: make install
|
||||
- run: make smoke-test
|
||||
|
||||
workflows:
|
||||
version: 2
|
||||
build:
|
||||
jobs:
|
||||
- build-python36
|
||||
#- build-python37
|
||||
#- build-python38 # no tensorflow for python 3.8
|
||||
|
|
@ -1,6 +0,0 @@
|
|||
tests
|
||||
dist
|
||||
build
|
||||
env*
|
||||
*.egg-info
|
||||
models_eynollah*
|
||||
44
.github/workflows/build-docker.yml
vendored
44
.github/workflows/build-docker.yml
vendored
|
|
@ -1,44 +0,0 @@
|
|||
name: CD
|
||||
|
||||
on:
|
||||
push:
|
||||
branches: [ "main" ]
|
||||
workflow_dispatch: # run manually
|
||||
|
||||
jobs:
|
||||
|
||||
build:
|
||||
runs-on: ubuntu-latest
|
||||
permissions:
|
||||
packages: write
|
||||
contents: read
|
||||
steps:
|
||||
- name: Checkout
|
||||
uses: actions/checkout@v4
|
||||
with:
|
||||
# we need tags for docker version tagging
|
||||
fetch-tags: true
|
||||
fetch-depth: 0
|
||||
- # Activate cache export feature to reduce build time of images
|
||||
name: Set up Docker Buildx
|
||||
uses: docker/setup-buildx-action@v3
|
||||
- name: Login to GitHub Container Registry
|
||||
uses: docker/login-action@v3
|
||||
with:
|
||||
registry: ghcr.io
|
||||
username: ${{ github.actor }}
|
||||
password: ${{ secrets.GITHUB_TOKEN }}
|
||||
- name: Log in to Docker Hub
|
||||
uses: docker/login-action@v3
|
||||
with:
|
||||
username: ${{ secrets.DOCKERIO_USERNAME }}
|
||||
password: ${{ secrets.DOCKERIO_PASSWORD }}
|
||||
- name: Build the Docker image
|
||||
# build both tags at the same time
|
||||
run: make docker DOCKER_TAG="docker.io/ocrd/eynollah ghcr.io/qurator-spk/eynollah"
|
||||
- name: Test the Docker image
|
||||
run: docker run --rm ocrd/eynollah ocrd-eynollah-segment -h
|
||||
- name: Push to Dockerhub
|
||||
run: docker push docker.io/ocrd/eynollah
|
||||
- name: Push to Github Container Registry
|
||||
run: docker push ghcr.io/qurator-spk/eynollah
|
||||
24
.github/workflows/pypi.yml
vendored
24
.github/workflows/pypi.yml
vendored
|
|
@ -1,24 +0,0 @@
|
|||
name: PyPI CD
|
||||
|
||||
on:
|
||||
release:
|
||||
types: [published]
|
||||
workflow_dispatch:
|
||||
|
||||
jobs:
|
||||
pypi-publish:
|
||||
name: upload release to PyPI
|
||||
runs-on: ubuntu-latest
|
||||
permissions:
|
||||
# IMPORTANT: this permission is mandatory for Trusted Publishing
|
||||
id-token: write
|
||||
steps:
|
||||
- uses: actions/checkout@v4
|
||||
- name: Set up Python
|
||||
uses: actions/setup-python@v5
|
||||
- name: Build package
|
||||
run: make build
|
||||
- name: Publish package distributions to PyPI
|
||||
uses: pypa/gh-action-pypi-publish@release/v1
|
||||
with:
|
||||
verbose: true
|
||||
90
.github/workflows/test-eynollah.yml
vendored
90
.github/workflows/test-eynollah.yml
vendored
|
|
@ -1,9 +1,9 @@
|
|||
# This workflow will install Python dependencies, run tests and lint with a variety of Python versions
|
||||
# For more information see: https://help.github.com/actions/language-and-framework-guides/using-python-with-github-actions
|
||||
|
||||
name: Test
|
||||
name: Python package
|
||||
|
||||
on: [push]
|
||||
on: [push, pull_request]
|
||||
|
||||
jobs:
|
||||
build:
|
||||
|
|
@ -11,90 +11,26 @@ jobs:
|
|||
runs-on: ubuntu-latest
|
||||
strategy:
|
||||
matrix:
|
||||
python-version: ['3.8', '3.9', '3.10', '3.11']
|
||||
python-version: ['3.6'] # '3.7'
|
||||
|
||||
steps:
|
||||
- name: clean up
|
||||
run: |
|
||||
df -h
|
||||
sudo rm -rf /usr/share/dotnet
|
||||
sudo rm -rf /usr/local/lib/android
|
||||
sudo rm -rf /opt/ghc
|
||||
sudo rm -rf "/usr/local/share/boost"
|
||||
sudo rm -rf "$AGENT_TOOLSDIRECTORY"
|
||||
df -h
|
||||
- uses: actions/checkout@v4
|
||||
|
||||
# - name: Lint with ruff
|
||||
# uses: astral-sh/ruff-action@v3
|
||||
# with:
|
||||
# src: "./src"
|
||||
|
||||
- name: Try to restore models_eynollah
|
||||
uses: actions/cache/restore@v4
|
||||
id: all_model_cache
|
||||
- uses: actions/checkout@v2
|
||||
- uses: actions/cache@v2
|
||||
id: model_cache
|
||||
with:
|
||||
path: models_eynollah
|
||||
key: models_eynollah-${{ hashFiles('src/eynollah/model_zoo/default_specs.py') }}
|
||||
|
||||
key: ${{ runner.os }}-models
|
||||
- name: Download models
|
||||
if: steps.all_model_cache.outputs.cache-hit != 'true'
|
||||
run: |
|
||||
make models
|
||||
ls -la models_eynollah
|
||||
|
||||
- uses: actions/cache/save@v4
|
||||
if: steps.all_model_cache.outputs.cache-hit != 'true'
|
||||
with:
|
||||
path: models_eynollah
|
||||
key: models_eynollah-${{ hashFiles('src/eynollah/model_zoo/default_specs.py') }}
|
||||
|
||||
if: steps.model_cache.outputs.cache-hit != 'true'
|
||||
run: make models
|
||||
- name: Set up Python ${{ matrix.python-version }}
|
||||
uses: actions/setup-python@v5
|
||||
uses: actions/setup-python@v2
|
||||
with:
|
||||
python-version: ${{ matrix.python-version }}
|
||||
|
||||
# - uses: actions/cache@v4
|
||||
# with:
|
||||
# path: |
|
||||
# path/to/dependencies
|
||||
# some/other/dependencies
|
||||
# key: ${{ runner.os }}-${{ hashFiles('**/lockfiles') }}
|
||||
|
||||
- name: Install dependencies
|
||||
run: |
|
||||
python -m pip install --upgrade pip
|
||||
make install-dev EXTRAS=OCR,plotting
|
||||
make deps-test EXTRAS=OCR,plotting
|
||||
|
||||
pip install .
|
||||
pip install -r requirements-test.txt
|
||||
- name: Test with pytest
|
||||
run: make coverage PYTEST_ARGS="-vv --junitxml=pytest.xml"
|
||||
|
||||
- name: Get coverage results
|
||||
run: |
|
||||
coverage report --format=markdown >> $GITHUB_STEP_SUMMARY
|
||||
coverage html
|
||||
coverage json
|
||||
coverage xml
|
||||
|
||||
- name: Store coverage results
|
||||
uses: actions/upload-artifact@v4
|
||||
with:
|
||||
name: coverage-report_${{ matrix.python-version }}
|
||||
path: |
|
||||
htmlcov
|
||||
pytest.xml
|
||||
coverage.xml
|
||||
coverage.json
|
||||
|
||||
- name: Upload coverage results
|
||||
uses: codecov/codecov-action@v4
|
||||
with:
|
||||
files: coverage.xml
|
||||
fail_ci_if_error: false
|
||||
|
||||
- name: Test standalone CLI
|
||||
run: make smoke-test
|
||||
|
||||
- name: Test OCR-D CLI
|
||||
run: make ocrd-test
|
||||
run: make test
|
||||
|
|
|
|||
7
.gitignore
vendored
7
.gitignore
vendored
|
|
@ -2,13 +2,6 @@
|
|||
__pycache__
|
||||
sbb_newspapers_org_image/pylint.log
|
||||
models_eynollah*
|
||||
models_ocr*
|
||||
models_layout*
|
||||
default-2021-03-09
|
||||
output.html
|
||||
/build
|
||||
/dist
|
||||
*.tif
|
||||
*.sw?
|
||||
TAGS
|
||||
uv.lock
|
||||
|
|
|
|||
318
CHANGELOG.md
318
CHANGELOG.md
|
|
@ -5,309 +5,7 @@ Versioned according to [Semantic Versioning](http://semver.org/).
|
|||
|
||||
## Unreleased
|
||||
|
||||
## [0.8.0] - 2026-05-11
|
||||
|
||||
* Optimize model performance
|
||||
* `multiprocessing.SpawnProcess` predictor wrapper for models to have commmunication with Tensorflow in a separate subprocess in a task queue with parallel jobs configurable via `--num-jobs` and maximum number of failed jobs via `--halt-fail`
|
||||
* Keep batch size low enough for processing fitting into common 8GB GPU (with model-dependent batch resizing prepared but not yet active)
|
||||
* GPU device can be selected manually with `--device`
|
||||
* Handle image resizing and tiling in GPU as much as possible to avoid overhead of switching between GPU and CPU
|
||||
* jit-compile and precompile models where possible (non-autosized, non-patched Keras models)
|
||||
* Fix bugs and homogenize internal labels related to differing labels for early layout and different stages of full layout detection
|
||||
* Replace `Lambda` layers with `ZeroPadding2D`, improving size and optimizability of models for `eynollah layout`
|
||||
* Improved training
|
||||
* Use connected components for loss function
|
||||
* Integrate with Tensorboard to observe model training progress, including plots and visualizing intermediate evaluation results
|
||||
* Simplified model usage
|
||||
* Models can be overridden individually, so any model trained with `eynollah-training` can replace any model in the [distributions on zenodo](https://zenodo.org/records/17727267)
|
||||
* `--model` is a CLI option of the `eynollah` root CLI now and should point to the same directory for all subcommands
|
||||
* Improved reading order detection heuristics
|
||||
* Improved drop capital, marginalia and column detection
|
||||
* Fixing bugs in polygon handling and image operations
|
||||
* No more self-intersecting polygons
|
||||
* Correct rotation implementation, enlarging/shrinking canvas as necessary
|
||||
* Use actual area of a polygon instead of length of polygon path or first candidate for comparisons
|
||||
* Improved PAGE-XML serialization
|
||||
* Annotate column classifier result in `/PcGts/Page/@custom` (Transkribus convention) and `/PcGts/Metadata/Comment` (QURATOR convention)
|
||||
* Annotate page skew in `/PcGts/Page/@orientation`
|
||||
* Calculate and annotate confidences as `Coords/@conf` for regions, lines, images and tables
|
||||
* Massive refactoring and code quality improvement
|
||||
* deduplication, idiomatic python, clean parallel processing, class reuse, consistent and meaningful naming
|
||||
|
||||
|
||||
**NOTE** We are aware of a possible issue with regards to the cropping of images. It appears that we have not consistenly cropped images for training. This can lead to suboptimal results for cropped images. If you experience quality issues with the `eynollah layout`, try setting the `-ipe/--ignore_page_extraction` option to skip the builtin cropping. We will rectify this in the next trainings.
|
||||
|
||||
## [0.7.0] - 2026-01-30
|
||||
|
||||
Added:
|
||||
|
||||
* "Model zoo", central place to describe and load models, #207
|
||||
* Training code for the CNN/RNN OCR model
|
||||
|
||||
Changed:
|
||||
|
||||
* Lint training code, #204
|
||||
* Update documentation: README, pyproject.toml metadata, guides in `docs/`, #209
|
||||
|
||||
|
||||
## [0.6.0] - 2025-10-17
|
||||
|
||||
Added:
|
||||
|
||||
* `eynollah-training` CLI and docs for training the models, #187, #193, https://github.com/qurator-spk/sbb_pixelwise_segmentation/tree/unifying-training-models
|
||||
|
||||
Fixed:
|
||||
|
||||
* `join_polygons` always returning Polygon, not MultiPolygon, #203
|
||||
|
||||
## [0.6.0rc2] - 2025-10-14
|
||||
|
||||
Fixed:
|
||||
|
||||
* Prevent OOM GPU error by avoiding loading the `region_fl` model, #199
|
||||
* XML output: encoding should be `utf-8`, not `utf8`, #196, #197
|
||||
|
||||
## [0.6.0rc1] - 2025-10-10
|
||||
|
||||
Fixed:
|
||||
|
||||
* continue processing when no columns detected but text regions exist
|
||||
* convert marginalia to main text if no main text is present
|
||||
* reset deskewing angle to 0° when text covers <30% image area and detected angle >45°
|
||||
* :fire: polygons: avoid invalid paths (use `Polygon.buffer()` instead of dilation etc.)
|
||||
* `return_boxes_of_images_by_order_of_reading_new`: avoid Numpy.dtype mismatch, simplify
|
||||
* `return_boxes_of_images_by_order_of_reading_new`: log any exceptions instead of ignoring
|
||||
* `filter_contours_without_textline_inside`: avoid removing from duplicate lists twice
|
||||
* `get_marginals`: exit early if no peaks found to avoid spurious overlap mask
|
||||
* `get_smallest_skew`: after shifting search range of rotation angle, use overall best result
|
||||
* Dockerfile: fix CUDA installation (cuDNN contested between Torch and TF due to extra OCR)
|
||||
* OCR: re-instate missing methods and fix `utils_ocr` function calls
|
||||
* mbreorder/enhancement CLIs: missing imports
|
||||
* :fire: writer: `SeparatorRegion` needs `SeparatorRegionType` (not `ImageRegionType`), f458e3e
|
||||
* tests: switch from `pytest-subtests` to `parametrize` so we can use `pytest-isolate`
|
||||
(so CUDA memory gets freed between tests if running on GPU)
|
||||
|
||||
Added:
|
||||
* :fire: `layout` CLI: new option `--model_version` to override default choices
|
||||
* test coverage for OCR options in `layout`
|
||||
* test coverage for table detection in `layout`
|
||||
* CI linting with ruff
|
||||
|
||||
Changed:
|
||||
|
||||
* polygons: slightly widen for regions and lines, increase for separators
|
||||
* various refactorings, some code style and identifier improvements
|
||||
* deskewing/multiprocessing: switch back to ProcessPoolExecutor (faster),
|
||||
but use shared memory if necessary, and switch back from `loky` to stdlib,
|
||||
and shutdown in `del()` instead of `atexit`
|
||||
* :fire: OCR: switch CNN-RNN model to `20250930` version compatible with TF 2.12 on CPU, too
|
||||
* OCR: allow running `-tr` without `-fl`, too
|
||||
* :fire: writer: use `@type='heading'` instead of `'header'` for headings
|
||||
* :fire: performance gains via refactoring (simplification, less copy-code, vectorization,
|
||||
avoiding unused calculations, avoiding unnecessary 3-channel image operations)
|
||||
* :fire: heuristic reading order detection: many improvements
|
||||
- contour vs splitter box matching:
|
||||
* contour must be contained in box exactly instead of heuristics
|
||||
* make fallback center matching, center must be contained in box
|
||||
- original vs deskewed contour matching:
|
||||
* same min-area filter on both sides
|
||||
* similar area score in addition to center proximity
|
||||
* avoid duplicate and missing mappings by allowing N:M
|
||||
matches and splitting+joining where necessary
|
||||
* CI: update+improve model caching
|
||||
|
||||
|
||||
## [0.5.0] - 2025-09-26
|
||||
|
||||
Fixed:
|
||||
|
||||
* restoring the contour in the original image caused an error due to an empty tuple, #154
|
||||
* removed NumPy warnings calculating sigma, mean, (fixed issue #158)
|
||||
* fixed bug in `separate_lines.py`, #124
|
||||
* Drop capitals are now handled separately from their corresponding textline
|
||||
* Marginals are now divided into left and right. Their reading order is written first for left marginals, then for right marginals, and within each side from top to bottom
|
||||
* Added a new page extraction model. Instead of bounding boxes, it outputs page contours in the XML file, improving results for skewed pages
|
||||
* Improved reading order for cases where a textline is segmented into multiple smaller textlines
|
||||
|
||||
Changed
|
||||
|
||||
* CLIs: read only allowed filename suffixes (image or XML) with `--dir_in`
|
||||
* CLIs: make all output option required, and `-i` / `-di` required but mutually exclusive
|
||||
* ocr CLI: drop redundant `-brb` in favour of just `-dib`
|
||||
* APIs: move all input/output path options from class (kwarg and attribute) ro `run` kwarg
|
||||
* layout textlines: polygonal also without `-cl`
|
||||
|
||||
Added:
|
||||
|
||||
* `eynollah machine-based-reading-order` CLI to run reading order detection, #175
|
||||
* `eynollah enhancement` CLI to run image enhancement, #175
|
||||
* Improved models for page extraction and reading order detection, #175
|
||||
* For the lightweight version (layout and textline detection), thresholds are now assigned to the artificial class. Users can apply these thresholds to improve detection of isolated textlines and regions. To counteract the drawback of thresholding, the skeleton of the artificial class is used to keep lines as thin as possible (resolved issues #163 and #161)
|
||||
* Added and integrated a trained CNN-RNN OCR models
|
||||
* Added and integrated a trained TrOCR model
|
||||
* Improved OCR detection to support vertical and curved textlines
|
||||
* Introduced a new machine-based reading order model with rotation augmentation
|
||||
* Optimized reading order speed by clustering text regions that belong to the same block, maintaining top-to-bottom order
|
||||
* Implemented text merging across textlines based on hyphenation when a line ends with a hyphen
|
||||
* Integrated image enhancement as a separate use case
|
||||
* Added reading order functionality on the layout level as a separate use case
|
||||
* CNN-RNN OCR models provide confidence scores for predictions
|
||||
* Added OCR visualization: predicted OCR can be overlaid on an image of the same size as the input
|
||||
* Introduced a threshold value for CNN-RNN OCR models, allowing users to filter out low-confidence textline predictions
|
||||
* For OCR, users can specify a single model by name instead of always using the default model
|
||||
* Under the OCR use case, if Ground Truth XMLs and images are available, textline image and corresponding text extraction can now be performed
|
||||
|
||||
Merged PRs:
|
||||
|
||||
* better machine based reading order + layout and textline + ocr by @vahidrezanezhad in https://github.com/qurator-spk/eynollah/pull/175
|
||||
* CI: pypi by @kba in https://github.com/qurator-spk/eynollah/pull/154
|
||||
* CI: Use most recent actions/setup-python@v5 by @kba in https://github.com/qurator-spk/eynollah/pull/157
|
||||
* update docker by @bertsky in https://github.com/qurator-spk/eynollah/pull/159
|
||||
* Ocrd fixes by @kba in https://github.com/qurator-spk/eynollah/pull/167
|
||||
* Updating readme for eynollah use cases cli by @kba in https://github.com/qurator-spk/eynollah/pull/166
|
||||
* OCR-D processor: expose reading_order_machine_based by @bertsky in https://github.com/qurator-spk/eynollah/pull/171
|
||||
* prepare release v0.5.0: fix logging by @bertsky in https://github.com/qurator-spk/eynollah/pull/180
|
||||
* mb_ro_on_layout: remove copy-pasta code not actually used by @kba in https://github.com/qurator-spk/eynollah/pull/181
|
||||
* prepare release v0.5.0: improve CLI docstring, refactor I/O path options from class to run kwargs, increase test coverage @bertsky in #182
|
||||
* prepare release v0.5.0: fix for OCR doit subtest by @bertsky in https://github.com/qurator-spk/eynollah/pull/183
|
||||
* Prepare release v0.5.0 by @kba in https://github.com/qurator-spk/eynollah/pull/178
|
||||
* updating eynollah README, how to use it for use cases by @vahidrezanezhad in https://github.com/qurator-spk/eynollah/pull/156
|
||||
* add feedback to command line interface by @michalbubula in https://github.com/qurator-spk/eynollah/pull/170
|
||||
|
||||
## [0.4.0] - 2025-04-07
|
||||
|
||||
Fixed:
|
||||
|
||||
* allow empty imports for optional dependencies
|
||||
* avoid Numpy warnings (empty slices etc.)
|
||||
* remove deprecated Numpy types
|
||||
* binarization CLI: make `dir_in` usable again
|
||||
|
||||
Added:
|
||||
|
||||
* Continuous Deployment via Dockerhub and GHCR
|
||||
* CI: also test CLIs and OCR-D
|
||||
* CI: measure code coverage, annotate+upload reports
|
||||
* smoke-test: also check results
|
||||
* smoke-test: also test sbb-binarize
|
||||
* ocrd-test: analog for OCR-D CLI (segment and binarize)
|
||||
* pytest: add asserts, extend coverage, use subtests for various options
|
||||
* pytest: also add binarization
|
||||
* pytest: add `dir_in` mode (segment and binarize)
|
||||
* make install: control optional dependencies via `EXTRAS` variable
|
||||
* OCR-D: expose and describe recently added parameters:
|
||||
- `ignore_page_extraction`
|
||||
- `allow_enhancement`
|
||||
- `textline_light`
|
||||
- `right_to_left`
|
||||
* OCR-D: :fire: integrate ocrd-sbb-binarize
|
||||
* add detection confidence in `TextRegion/Coords/@conf`
|
||||
(but only in light version and not for marginalia)
|
||||
|
||||
Changed:
|
||||
|
||||
* Docker build: simplify, w/ `OCR`, conform to OCR-D spec
|
||||
* OCR-D: :fire: migrate to core v3
|
||||
- initialize+setup only once
|
||||
- restrict number of parallel page workers to 1
|
||||
(conflicts with existing multiprocessing; TF parts not mp-compatible)
|
||||
- do query maximally annotated page image
|
||||
(but filtering existing binarization/cropping/deskewing),
|
||||
rebase (as new `@imageFilename`) if necessary
|
||||
- add behavioural docstring
|
||||
|
||||
* :fire: refactor `Eynollah` API:
|
||||
- no more data (kw)args at init,
|
||||
but kwargs `dir_in` / `image_filename` for `run()`
|
||||
- no more data attributes, but function kwargs
|
||||
(`pcgts`, `image_filename`, `image_pil`, `dir_in`, `override_dpi`)
|
||||
- remove redundant TF session/model loaders
|
||||
(only load once during init)
|
||||
- factor `run_single()` out of `run()` (loop body),
|
||||
expose for independent calls (like OCR-D)
|
||||
- expose `cache_images()`, add `dpi` kwarg, set `self._imgs`
|
||||
- single-image mode writes PAGE file result
|
||||
(just as directory mode does)
|
||||
|
||||
* CLI: assertions (instead of print+exit) for options checks
|
||||
* light mode: fine-tune ratio to better detect a region as header
|
||||
|
||||
## [0.3.1] - 2024-08-27
|
||||
|
||||
Fixed:
|
||||
|
||||
* regression in OCR-D processor, #106
|
||||
* Expected Ptrcv::UMat for argument 'contour', #110
|
||||
* Memory usage explosion with very narrow images (e.g. book spine), #67
|
||||
|
||||
## [0.3.0] - 2023-05-13
|
||||
|
||||
Changed:
|
||||
|
||||
* Eynollah light integration, #86
|
||||
* use PEP420 style qurator namespace, #97
|
||||
* set_memory_growth to all GPU devices alike, #100
|
||||
|
||||
Fixed:
|
||||
|
||||
* PAGE-XML coordinates can have self-intersections, #20
|
||||
* reading order representation (XML order vs index), #22
|
||||
* allow cropping separately, #26
|
||||
* Order of regions, #51
|
||||
* error while running inference, #75
|
||||
* Eynollah crashes while processing image, #77
|
||||
* ValueError: bad marshal data, #87
|
||||
* contour extraction: inhomogeneous shape, #92
|
||||
* Confusing model dir variables, #93
|
||||
* New release?, #96
|
||||
|
||||
## [0.2.0] - 2023-03-24
|
||||
|
||||
Changed:
|
||||
|
||||
* Convert default model from HDFS to TF SavedModel, #91
|
||||
|
||||
Added:
|
||||
|
||||
* parmeter `tables` to toggle table detectino, #91
|
||||
* default model described in ocrd-tool.json, #91
|
||||
|
||||
## [0.1.0] - 2023-03-22
|
||||
|
||||
Fixed:
|
||||
|
||||
* Do not produce spurious `TextEquiv`, #68
|
||||
* Less spammy logging, #64, #65, #71
|
||||
|
||||
Changed:
|
||||
|
||||
* Upgrade to tensorflow 2.4.0, #74
|
||||
* Improved README
|
||||
* CI: test for python 3.7+, #90
|
||||
|
||||
## [0.0.11] - 2022-02-02
|
||||
|
||||
Fixed:
|
||||
|
||||
* `models` parameter should have `content-type`, #61, OCR-D/core#777
|
||||
|
||||
## [0.0.10] - 2021-09-27
|
||||
|
||||
Fixed:
|
||||
|
||||
* call to `uild_pagexml_no_full_layout` for empty pages, #52
|
||||
|
||||
## [0.0.9] - 2021-08-16
|
||||
|
||||
Added:
|
||||
|
||||
* Table detection, #48
|
||||
|
||||
Fixed:
|
||||
|
||||
* Catch exception, #47
|
||||
|
||||
## [0.0.8] - 2021-07-27
|
||||
## [0.0.7] - 2021-07-27
|
||||
|
||||
Fixed:
|
||||
|
||||
|
|
@ -352,20 +50,6 @@ Fixed:
|
|||
Initial release
|
||||
|
||||
<!-- link-labels -->
|
||||
[0.8.0]: ../../compare/v0.8.0...v0.7.0
|
||||
[0.7.0]: ../../compare/v0.7.0...v0.6.0
|
||||
[0.6.0]: ../../compare/v0.6.0...v0.6.0rc2
|
||||
[0.6.0rc2]: ../../compare/v0.6.0rc2...v0.6.0rc1
|
||||
[0.6.0rc1]: ../../compare/v0.6.0rc1...v0.5.0
|
||||
[0.5.0]: ../../compare/v0.5.0...v0.4.0
|
||||
[0.4.0]: ../../compare/v0.4.0...v0.3.1
|
||||
[0.3.1]: ../../compare/v0.3.1...v0.3.0
|
||||
[0.3.0]: ../../compare/v0.3.0...v0.2.0
|
||||
[0.2.0]: ../../compare/v0.2.0...v0.1.0
|
||||
[0.1.0]: ../../compare/v0.1.0...v0.0.11
|
||||
[0.0.11]: ../../compare/v0.0.11...v0.0.10
|
||||
[0.0.10]: ../../compare/v0.0.10...v0.0.9
|
||||
[0.0.9]: ../../compare/v0.0.9...v0.0.8
|
||||
[0.0.8]: ../../compare/v0.0.8...v0.0.7
|
||||
[0.0.7]: ../../compare/v0.0.7...v0.0.6
|
||||
[0.0.6]: ../../compare/v0.0.6...v0.0.5
|
||||
|
|
|
|||
49
Dockerfile
49
Dockerfile
|
|
@ -1,49 +0,0 @@
|
|||
ARG DOCKER_BASE_IMAGE
|
||||
FROM $DOCKER_BASE_IMAGE
|
||||
|
||||
ARG VCS_REF
|
||||
ARG BUILD_DATE
|
||||
LABEL \
|
||||
maintainer="https://ocr-d.de/en/contact" \
|
||||
org.label-schema.vcs-ref=$VCS_REF \
|
||||
org.label-schema.vcs-url="https://github.com/qurator-spk/eynollah" \
|
||||
org.label-schema.build-date=$BUILD_DATE \
|
||||
org.opencontainers.image.vendor="DFG-Funded Initiative for Optical Character Recognition Development" \
|
||||
org.opencontainers.image.title="Eynollah" \
|
||||
org.opencontainers.image.description="" \
|
||||
org.opencontainers.image.source="https://github.com/qurator-spk/eynollah" \
|
||||
org.opencontainers.image.documentation="https://github.com/qurator-spk/eynollah/blob/${VCS_REF}/README.md" \
|
||||
org.opencontainers.image.revision=$VCS_REF \
|
||||
org.opencontainers.image.created=$BUILD_DATE \
|
||||
org.opencontainers.image.base.name=ocrd/core-cuda-tf2
|
||||
|
||||
ENV DEBIAN_FRONTEND=noninteractive
|
||||
# set proper locales
|
||||
ENV PYTHONIOENCODING=utf8
|
||||
ENV LANG=C.UTF-8
|
||||
ENV LC_ALL=C.UTF-8
|
||||
|
||||
# avoid HOME/.local/share (hard to predict USER here)
|
||||
# so let XDG_DATA_HOME coincide with fixed system location
|
||||
# (can still be overridden by derived stages)
|
||||
ENV XDG_DATA_HOME /usr/local/share
|
||||
# avoid the need for an extra volume for persistent resource user db
|
||||
# (i.e. XDG_CONFIG_HOME/ocrd/resources.yml)
|
||||
ENV XDG_CONFIG_HOME /usr/local/share/ocrd-resources
|
||||
|
||||
WORKDIR /build/eynollah
|
||||
COPY . .
|
||||
COPY ocrd-tool.json .
|
||||
# prepackage ocrd-tool.json as ocrd-all-tool.json
|
||||
RUN ocrd ocrd-tool ocrd-tool.json dump-tools > $(dirname $(ocrd bashlib filename))/ocrd-all-tool.json
|
||||
# prepackage ocrd-all-module-dir.json
|
||||
RUN ocrd ocrd-tool ocrd-tool.json dump-module-dirs > $(dirname $(ocrd bashlib filename))/ocrd-all-module-dir.json
|
||||
# install everything and reduce image size
|
||||
RUN make install EXTRAS=OCR && rm -rf /build/eynollah
|
||||
# fixup for broken cuDNN installation (Torch pulls in 8.5.0, which is incompatible with Tensorflow)
|
||||
RUN pip install nvidia-cudnn-cu11==8.6.0.163
|
||||
# smoke test
|
||||
RUN eynollah --help
|
||||
|
||||
WORKDIR /data
|
||||
VOLUME /data
|
||||
124
Makefile
124
Makefile
|
|
@ -1,23 +1,5 @@
|
|||
PYTHON ?= python3
|
||||
PIP ?= pip3
|
||||
EXTRAS ?=
|
||||
|
||||
DOCKER_BASE_IMAGE ?= docker.io/ocrd/core-cuda-tf2:v3.13.0
|
||||
DOCKER_TAG ?= ocrd/eynollah
|
||||
DOCKER ?= docker
|
||||
WGET = wget -O
|
||||
|
||||
#SEG_MODEL := https://qurator-data.de/eynollah/2021-04-25/models_eynollah.tar.gz
|
||||
#SEG_MODEL := https://qurator-data.de/eynollah/2022-04-05/models_eynollah_renamed.tar.gz
|
||||
# SEG_MODEL := https://qurator-data.de/eynollah/2022-04-05/models_eynollah.tar.gz
|
||||
#SEG_MODEL := https://github.com/qurator-spk/eynollah/releases/download/v0.3.0/models_eynollah.tar.gz
|
||||
#SEG_MODEL := https://github.com/qurator-spk/eynollah/releases/download/v0.3.1/models_eynollah.tar.gz
|
||||
#SEG_MODEL := https://zenodo.org/records/17194824/files/models_layout_v0_5_0.tar.gz?download=1
|
||||
EYNOLLAH_MODELS_URL := https://zenodo.org/records/17727267/files/models_all_v0_8_0.zip
|
||||
EYNOLLAH_MODELS_ZIP = $(notdir $(EYNOLLAH_MODELS_URL))
|
||||
EYNOLLAH_MODELS_DIR = $(EYNOLLAH_MODELS_ZIP:%.zip=%)
|
||||
|
||||
PYTEST_ARGS ?= -vv --isolate
|
||||
EYNOLLAH_MODELS ?= $(PWD)/models_eynollah
|
||||
export EYNOLLAH_MODELS
|
||||
|
||||
# BEGIN-EVAL makefile-parser --make-help Makefile
|
||||
|
||||
|
|
@ -25,113 +7,37 @@ help:
|
|||
@echo ""
|
||||
@echo " Targets"
|
||||
@echo ""
|
||||
@echo " docker Build Docker image"
|
||||
@echo " build Build Python source and binary distribution"
|
||||
@echo " install Install package with pip"
|
||||
@echo " models Download and extract models to $(PWD)/models_eynollah"
|
||||
@echo " install Install with pip"
|
||||
@echo " install-dev Install editable with pip"
|
||||
@echo " deps-test Install test dependencies with pip"
|
||||
@echo " models Download and extract models to $(CURDIR):"
|
||||
@echo " $(EYNOLLAH_MODELS_DIR)"
|
||||
@echo " smoke-test Run simple CLI check"
|
||||
@echo " ocrd-test Run OCR-D CLI check"
|
||||
@echo " test Run unit tests"
|
||||
@echo ""
|
||||
@echo " Variables"
|
||||
@echo " EXTRAS comma-separated list of features (like 'OCR,plotting') for 'install' [$(EXTRAS)]"
|
||||
@echo " DOCKER_TAG Docker image tag for 'docker' [$(DOCKER_TAG)]"
|
||||
@echo " PYTEST_ARGS pytest args for 'test' (Set to '-s' to see log output during test execution, '-vv' to see individual tests. [$(PYTEST_ARGS)]"
|
||||
@echo " ALL_MODELS URL of archive of all models [$(ALL_MODELS)]"
|
||||
@echo ""
|
||||
|
||||
# END-EVAL
|
||||
|
||||
# Download and extract models to $(PWD)/models_layout_v0_6_0
|
||||
models: $(EYNOLLAH_MODELS_DIR)
|
||||
|
||||
# do not download these files if we already have the directories
|
||||
.INTERMEDIATE: $(EYNOLLAH_MODELS_ZIP)
|
||||
# Download and extract models to $(PWD)/models_eynollah
|
||||
models: models_eynollah
|
||||
|
||||
$(EYNOLLAH_MODELS_ZIP):
|
||||
$(WGET) $@ $(EYNOLLAH_MODELS_URL)
|
||||
models_eynollah: models_eynollah.tar.gz
|
||||
tar xf models_eynollah.tar.gz
|
||||
|
||||
$(EYNOLLAH_MODELS_DIR): $(EYNOLLAH_MODELS_ZIP)
|
||||
unzip $<
|
||||
|
||||
build:
|
||||
$(PIP) install build
|
||||
$(PYTHON) -m build .
|
||||
models_eynollah.tar.gz:
|
||||
wget 'https://qurator-data.de/eynollah/models_eynollah.tar.gz'
|
||||
|
||||
# Install with pip
|
||||
install:
|
||||
$(PIP) install .$(and $(EXTRAS),[$(EXTRAS)])
|
||||
pip install .
|
||||
|
||||
# Install editable with pip
|
||||
install-dev:
|
||||
$(PIP) install -e .$(and $(EXTRAS),[$(EXTRAS)])
|
||||
pip install -e .
|
||||
|
||||
deps-test:
|
||||
$(PIP) install -r requirements-test.txt
|
||||
|
||||
smoke-test: TMPDIR != mktemp -d
|
||||
smoke-test: tests/resources/2files/kant_aufklaerung_1784_0020.tif
|
||||
# layout analysis:
|
||||
eynollah -m $(CURDIR) layout -i $< -o $(TMPDIR)
|
||||
fgrep -q http://schema.primaresearch.org/PAGE/gts/pagecontent/2019-07-15 $(TMPDIR)/$(basename $(<F)).xml
|
||||
fgrep -c -e TextRegion -e ImageRegion -e SeparatorRegion $(TMPDIR)/$(basename $(<F)).xml
|
||||
# layout, directory mode (skip one, add one):
|
||||
eynollah -m $(CURDIR) layout -di $(<D) -o $(TMPDIR)
|
||||
test -s $(TMPDIR)/euler_rechenkunst01_1738_0025.xml
|
||||
# mbreorder, directory mode (overwrite):
|
||||
eynollah -m $(CURDIR) machine-based-reading-order -di $(<D) -o $(TMPDIR)
|
||||
fgrep -q http://schema.primaresearch.org/PAGE/gts/pagecontent/2019-07-15 $(TMPDIR)/$(basename $(<F)).xml
|
||||
fgrep -c -e RegionRefIndexed $(TMPDIR)/$(basename $(<F)).xml
|
||||
# binarize:
|
||||
eynollah -m $(CURDIR) binarization -i $< -o $(TMPDIR)/$(<F)
|
||||
test -s $(TMPDIR)/$(<F)
|
||||
@set -x; test "$$(identify -format '%w %h' $<)" = "$$(identify -format '%w %h' $(TMPDIR)/$(<F))"
|
||||
# enhance:
|
||||
eynollah -m $(CURDIR) enhancement -sos -i $< -o $(TMPDIR) -O
|
||||
test -s $(TMPDIR)/$(<F)
|
||||
@set -x; test "$$(identify -format '%w %h' $<)" = "$$(identify -format '%w %h' $(TMPDIR)/$(<F))"
|
||||
$(RM) -r $(TMPDIR)
|
||||
|
||||
ocrd-test: export OCRD_MISSING_OUTPUT := ABORT
|
||||
ocrd-test: TMPDIR != mktemp -d
|
||||
ocrd-test: tests/resources/2files/kant_aufklaerung_1784_0020.tif
|
||||
cp $< $(TMPDIR)
|
||||
ocrd workspace -d $(TMPDIR) init
|
||||
ocrd workspace -d $(TMPDIR) add -G OCR-D-IMG -g PHYS_0020 -i OCR-D-IMG_0020 $(<F)
|
||||
ocrd-eynollah-segment -w $(TMPDIR) -I OCR-D-IMG -O OCR-D-SEG -P models $(CURDIR)
|
||||
result=$$(ocrd workspace -d $(TMPDIR) find -G OCR-D-SEG); \
|
||||
fgrep -q http://schema.primaresearch.org/PAGE/gts/pagecontent/2019-07-15 $(TMPDIR)/$$result && \
|
||||
fgrep -c -e TextRegion -e ImageRegion -e SeparatorRegion $(TMPDIR)/$$result
|
||||
ocrd-sbb-binarize -w $(TMPDIR) -I OCR-D-IMG -O OCR-D-BIN -P model $(CURDIR)
|
||||
ocrd-sbb-binarize -w $(TMPDIR) -I OCR-D-SEG -O OCR-D-SEG-BIN -P model $(CURDIR) -P operation_level region
|
||||
$(RM) -r $(TMPDIR)
|
||||
smoke-test:
|
||||
eynollah -i tests/resources/kant_aufklaerung_1784_0020.tif -o . -m $(PWD)/models_eynollah
|
||||
|
||||
# Run unit tests
|
||||
test: export EYNOLLAH_MODELS_DIR := $(CURDIR)
|
||||
test:
|
||||
$(PYTHON) -m pytest tests --durations=0 --continue-on-collection-errors $(PYTEST_ARGS)
|
||||
|
||||
coverage:
|
||||
coverage erase
|
||||
$(MAKE) test PYTHON="coverage run"
|
||||
coverage report -m
|
||||
|
||||
# Concatenate docker image names with either the git tag describing current commit or 'latest' and
|
||||
# merge list with "-t"
|
||||
empty :=
|
||||
space := $(empty) $(empty)
|
||||
GIT_TAG := $(strip $(shell git describe --tags | grep -x "v[0-9]\+\.[0-9]\+\.[0-9]\+"))
|
||||
DOCKER_TAGS = $(subst $(space),$(space)-t$(space),$(DOCKER_TAG:%=$(if $(GIT_TAG),%:$(GIT_TAG),%:latest)))
|
||||
|
||||
# Build docker image
|
||||
docker:
|
||||
$(DOCKER) build \
|
||||
--build-arg DOCKER_BASE_IMAGE=$(DOCKER_BASE_IMAGE) \
|
||||
--build-arg VCS_REF=$$(git rev-parse --short HEAD) \
|
||||
--build-arg BUILD_DATE=$$(date -u +"%Y-%m-%dT%H:%M:%SZ") \
|
||||
-t $(DOCKER_TAGS) .
|
||||
|
||||
.PHONY: models build install install-dev test smoke-test ocrd-test coverage docker help
|
||||
pytest tests
|
||||
|
|
|
|||
268
README.md
268
README.md
|
|
@ -1,212 +1,122 @@
|
|||
# Eynollah
|
||||
|
||||
> Document Layout Analysis, Binarization and OCR with Deep Learning and Heuristics
|
||||
|
||||
[](https://pypi.python.org/pypi/eynollah)
|
||||
[](https://pypi.org/project/eynollah/)
|
||||
[](https://github.com/qurator-spk/eynollah/actions/workflows/test-eynollah.yml)
|
||||
[](https://github.com/qurator-spk/eynollah/actions/workflows/build-docker.yml)
|
||||
[](https://opensource.org/license/apache-2-0/)
|
||||
[](https://doi.org/10.1145/3604951.3605513)
|
||||
> Document Layout Analysis
|
||||
|
||||
|
||||
|
||||
## Features
|
||||
* Document layout analysis using pixelwise segmentation models with support for 10 segmentation classes:
|
||||
* background, [page border](https://ocr-d.de/en/gt-guidelines/trans/lyRand.html), [text region](https://ocr-d.de/en/gt-guidelines/trans/lytextregion.html#textregionen__textregion_), [text line](https://ocr-d.de/en/gt-guidelines/pagexml/pagecontent_xsd_Complex_Type_pc_TextLineType.html), [header](https://ocr-d.de/en/gt-guidelines/trans/lyUeberschrift.html), [image](https://ocr-d.de/en/gt-guidelines/trans/lyBildbereiche.html), [separator](https://ocr-d.de/en/gt-guidelines/trans/lySeparatoren.html), [marginalia](https://ocr-d.de/en/gt-guidelines/trans/lyMarginalie.html), [initial](https://ocr-d.de/en/gt-guidelines/trans/lyInitiale.html), [table](https://ocr-d.de/en/gt-guidelines/trans/lyTabellen.html)
|
||||
* Textline segmentation to bounding boxes or polygons (contours) including for curved lines and vertical text
|
||||
* Document image binarization with pixelwise segmentation or hybrid CNN-Transformer models
|
||||
* Text recognition (OCR) with CNN-RNN or TrOCR models
|
||||
* Detection of reading order (left-to-right or right-to-left) using heuristics or trainable models
|
||||
* Output in [PAGE-XML](https://github.com/PRImA-Research-Lab/PAGE-XML)
|
||||
* [OCR-D](https://github.com/qurator-spk/eynollah#use-as-ocr-d-processor) interface
|
||||
## Introduction
|
||||
This tool performs document layout analysis (segmentation) from image data and returns the results as [PAGE-XML](https://github.com/PRImA-Research-Lab/PAGE-XML).
|
||||
|
||||
:warning: Development is focused on achieving the best quality of results for a wide variety of historical
|
||||
documents using a combination of multiple deep learning models and heuristics; therefore processing can be slow.
|
||||
It can currently detect the following layout classes/elements:
|
||||
* [Border](https://ocr-d.de/en/gt-guidelines/pagexml/pagecontent_xsd_Complex_Type_pc_BorderType.html)
|
||||
* [Textregion](https://ocr-d.de/en/gt-guidelines/pagexml/pagecontent_xsd_Complex_Type_pc_TextRegionType.html)
|
||||
* [Textline](https://ocr-d.de/en/gt-guidelines/pagexml/pagecontent_xsd_Complex_Type_pc_TextLineType.html)
|
||||
* [Image](https://ocr-d.de/en/gt-guidelines/pagexml/pagecontent_xsd_Complex_Type_pc_ImageRegionType.html)
|
||||
* [Separator](https://ocr-d.de/en/gt-guidelines/pagexml/pagecontent_xsd_Complex_Type_pc_SeparatorRegionType.html)
|
||||
* [Marginalia](https://ocr-d.de/en/gt-guidelines/trans/lyMarginalie.html)
|
||||
* [Initial (Drop Capital)](https://ocr-d.de/en/gt-guidelines/trans/lyInitiale.html)
|
||||
|
||||
In addition, the tool can be used to detect the _[ReadingOrder](https://ocr-d.de/en/gt-guidelines/trans/lyLeserichtung.html)_ of regions. The final goal is to feed the output to an OCR model.
|
||||
|
||||
The tool uses a combination of various models and heuristics (see flowchart below for the different stages and how they interact):
|
||||
* [Border detection](https://github.com/qurator-spk/eynollah#border-detection)
|
||||
* [Layout detection](https://github.com/qurator-spk/eynollah#layout-detection)
|
||||
* [Textline detection](https://github.com/qurator-spk/eynollah#textline-detection)
|
||||
* [Image enhancement](https://github.com/qurator-spk/eynollah#Image_enhancement)
|
||||
* [Scale classification](https://github.com/qurator-spk/eynollah#Scale_classification)
|
||||
* [Heuristic methods](https://https://github.com/qurator-spk/eynollah#heuristic-methods)
|
||||
|
||||
The first three stages are based on [pixel-wise segmentation](https://github.com/qurator-spk/sbb_pixelwise_segmentation).
|
||||
|
||||
|
||||
|
||||
## Border detection
|
||||
For the purpose of text recognition (OCR) and in order to avoid noise being introduced from texts outside the printspace, one first needs to detect the border of the printed frame. This is done by a binary pixel-wise-segmentation model trained on a dataset of 2,000 documents where about 1,200 of them come from the [dhSegment](https://github.com/dhlab-epfl/dhSegment/) project (you can download the dataset from [here](https://github.com/dhlab-epfl/dhSegment/releases/download/v0.2/pages.zip)) and the remainder having been annotated in SBB. For border detection, the model needs to be fed with the whole image at once rather than separated in patches.
|
||||
|
||||
## Layout detection
|
||||
As a next step, text regions need to be identified by means of layout detection. Again a pixel-wise segmentation model was trained on 131 labeled images from the SBB digital collections, including some data augmentation. Since the target of this tool are historical documents, we consider as main region types text regions, separators, images, tables and background - each with their own subclasses, e.g. in the case of text regions, subclasses like header/heading, drop capital, main body text etc. While it would be desirable to detect and classify each of these classes in a granular way, there are also limitations due to having a suitably large and balanced training set. Accordingly, the current version of this tool is focussed on the main region types background, text region, image and separator.
|
||||
|
||||
## Textline detection
|
||||
In a subsequent step, binary pixel-wise segmentation is used again to classify pixels in a document that constitute textlines. For textline segmentation, a model was initially trained on documents with only one column/block of text and some augmentation with regard to scaling. By fine-tuning the parameters also for multi-column documents, additional training data was produced that resulted in a much more robust textline detection model.
|
||||
|
||||
## Image enhancement
|
||||
This is an image to image model which input was low quality of an image and label was actually the original image. For this one we did not have any GT, so we decreased the quality of documents in SBB and then feed them into model.
|
||||
|
||||
## Scale classification
|
||||
This is simply an image classifier which classifies images based on their scales or better to say based on their number of columns.
|
||||
|
||||
## Heuristic methods
|
||||
Some heuristic methods are also employed to further improve the model predictions:
|
||||
* After border detection, the largest contour is determined by a bounding box, and the image cropped to these coordinates.
|
||||
* For text region detection, the image is scaled up to make it easier for the model to detect background space between text regions.
|
||||
* A minimum area is defined for text regions in relation to the overall image dimensions, so that very small regions that are noise can be filtered out.
|
||||
* Deskewing is applied on the text region level (due to regions having different degrees of skew) in order to improve the textline segmentation result.
|
||||
* After deskewing, a calculation of the pixel distribution on the X-axis allows the separation of textlines (foreground) and background pixels.
|
||||
* Finally, using the derived coordinates, bounding boxes are determined for each textline.
|
||||
|
||||
## Installation
|
||||
Python `3.8-3.11` with Tensorflow `<2.13` on Linux are currently supported.
|
||||
For (limited) GPU support the CUDA toolkit needs to be installed.
|
||||
A working config is CUDA `11.8` with cuDNN `8.6`.
|
||||
`pip install .` or
|
||||
|
||||
You can either install from PyPI
|
||||
`pip install . -e` for editable installation
|
||||
|
||||
```
|
||||
pip install eynollah
|
||||
```
|
||||
Alternatively, you can also use `make` with these targets:
|
||||
|
||||
or clone the repository, enter it and install (editable) with
|
||||
`make install` or
|
||||
|
||||
```
|
||||
git clone git@github.com:qurator-spk/eynollah.git
|
||||
cd eynollah; pip install -e .
|
||||
```
|
||||
`make install-dev` for editable installation
|
||||
|
||||
Alternatively, you can run `make install` or `make install-dev` for editable installation.
|
||||
### Models
|
||||
|
||||
To also install the dependencies for the OCR engines:
|
||||
In order to run this tool you also need trained models. You can download our pretrained models from [qurator-data.de](https://qurator-data.de/eynollah/).
|
||||
|
||||
```
|
||||
pip install "eynollah[OCR]"
|
||||
# or
|
||||
make install EXTRAS=OCR
|
||||
```
|
||||
|
||||
### Docker
|
||||
|
||||
Use
|
||||
|
||||
```
|
||||
docker pull ghcr.io/qurator-spk/eynollah:latest
|
||||
```
|
||||
|
||||
When using Eynollah with Docker, see [`docker.md`](https://github.com/qurator-spk/eynollah/tree/main/docs/docker.md).
|
||||
|
||||
## Models
|
||||
|
||||
Pretrained models can be downloaded from [Zenodo](https://zenodo.org/records/17727267) or [Hugging Face](https://huggingface.co/SBB?search_models=eynollah).
|
||||
|
||||
For model documentation and model cards, see [`models.md`](https://github.com/qurator-spk/eynollah/tree/main/docs/models.md).
|
||||
|
||||
## Training
|
||||
|
||||
To train your own model with Eynollah, see [`train.md`](https://github.com/qurator-spk/eynollah/tree/main/docs/train.md) and use the tools in the [`train`](https://github.com/qurator-spk/eynollah/tree/main/train) folder.
|
||||
Alternatively, running `make models` will download and extract models to `$(PWD)/models_eynollah`.
|
||||
|
||||
## Usage
|
||||
|
||||
Eynollah supports five use cases:
|
||||
1. [layout analysis (segmentation)](#layout-analysis),
|
||||
2. [binarization](#binarization),
|
||||
3. [image enhancement](#image-enhancement),
|
||||
4. [text recognition (OCR)](#ocr), and
|
||||
5. [reading order detection](#reading-order-detection).
|
||||
|
||||
Some example outputs can be found in [`examples.md`](https://github.com/qurator-spk/eynollah/tree/main/docs/examples.md).
|
||||
|
||||
### Layout Analysis
|
||||
|
||||
The layout analysis module is responsible for detecting layout elements, identifying text lines, and determining reading
|
||||
order using heuristic methods or a [pretrained model](https://github.com/qurator-spk/eynollah#machine-based-reading-order).
|
||||
|
||||
The command-line interface for layout analysis can be called like this:
|
||||
The basic command-line interface can be called like this:
|
||||
|
||||
```sh
|
||||
eynollah layout \
|
||||
-i <single image file> | -di <directory containing image files> \
|
||||
-o <output directory> \
|
||||
-m <directory containing model files> \
|
||||
[OPTIONS]
|
||||
eynollah \
|
||||
-i <image file name> \
|
||||
-o <directory to write output xml or enhanced image> \
|
||||
-m <directory of models> \
|
||||
-fl <if true, the tool will perform full layout analysis> \
|
||||
-ae <if true, the tool will resize and enhance the image and produce the resulting image as output> \
|
||||
-as <if true, the tool will check whether the document needs rescaling or not> \
|
||||
-cl <if true, the tool will extract the contours of curved textlines instead of rectangle bounding boxes> \
|
||||
-si <if a directory is given here, the tool will output image regions inside documents there>
|
||||
```
|
||||
|
||||
The following options can be used to further configure the processing:
|
||||
The tool does accept and works better on original images (RGB format) than binarized images.
|
||||
|
||||
| option | description |
|
||||
|-------------------|:--------------------------------------------------------------------------------------------|
|
||||
| `-fl` | full layout analysis including all steps and segmentation classes (recommended) |
|
||||
| `-tab` | apply table detection |
|
||||
| `-ae` | apply enhancement (the resulting image is saved to the output directory) |
|
||||
| `-as` | apply scaling |
|
||||
| `-cl` | apply contour detection for curved text lines instead of bounding boxes |
|
||||
| `-ib` | apply binarization (the resulting image is saved to the output directory) |
|
||||
| `-ep` | enable plotting (MUST always be used with `-sl`, `-sd`, `-sa`, `-si` or `-ae`) |
|
||||
| `-ho` | ignore headers for reading order dectection |
|
||||
| `-si <directory>` | save image regions detected to this directory |
|
||||
| `-sd <directory>` | save deskewed image to this directory |
|
||||
| `-sl <directory>` | save layout prediction as plot to this directory |
|
||||
| `-sp <directory>` | save cropped page image to this directory |
|
||||
| `-sa <directory>` | save all (plot, enhanced/binary image, layout) to this directory |
|
||||
| `-thart` | threshold of artifical class in the case of textline detection. The default value is 0.1 |
|
||||
| `-tharl` | threshold of artifical class in the case of layout detection. The default value is 0.1 |
|
||||
| `-ncu` | upper limit of columns in document image |
|
||||
| `-ncl` | lower limit of columns in document image |
|
||||
| `-slro` | skip layout detection and reading order |
|
||||
| `-romb` | apply machine based reading order detection |
|
||||
| `-ipe` | ignore page extraction |
|
||||
### `--full-layout` vs `--no-full-layout`
|
||||
|
||||
Here are the difference in elements detected depending on the `--full-layout`/`--no-full-layout` command line flags:
|
||||
|
||||
If no further option is set, the tool performs layout detection of main regions (background, text, images, separators
|
||||
and marginals).
|
||||
The best output quality is achieved when RGB images are used as input rather than greyscale or binarized images.
|
||||
| | `--full-layout` | `--no-full-layout` |
|
||||
| --- | --- | --- |
|
||||
| reading order | x | x |
|
||||
| header regions | x | - |
|
||||
| text regions | x | x |
|
||||
| text regions / text line | x | x |
|
||||
| drop-capitals | x | - |
|
||||
| marginals | x | x |
|
||||
| marginals / text line | x | x |
|
||||
| image region | x | x |
|
||||
|
||||
Additional documentation can be found in [`usage.md`](https://github.com/qurator-spk/eynollah/tree/main/docs/usage.md).
|
||||
### How to use
|
||||
|
||||
### Binarization
|
||||
First, this model makes use of up to 9 trained models which are responsible for different operations like size detection, column classification, image enhancement, page extraction, main layout detection, full layout detection and textline detection.That does not mean that all 9 models are always required for every document. Based on the document characteristics and parameters specified, different scenarios can be applied.
|
||||
|
||||
The binarization module performs document image binarization using pretrained pixelwise segmentation models.
|
||||
* If none of the parameters is set to `true`, the tool will perform a layout detection of main regions (background, text, images, separators and marginals). An advantage of this tool is that it tries to extract main text regions separately as much as possible.
|
||||
|
||||
The command-line interface for binarization can be called like this:
|
||||
* If you set `-ae` (**a**llow image **e**nhancement) parameter to `true`, the tool will first check the ppi (pixel-per-inch) of the image and when it is less than 300, the tool will resize it and only then image enhancement will occur. Image enhancement can also take place without this option, but by setting this option to `true`, the layout xml data (e.g. coordinates) will be based on the resized and enhanced image instead of the original image.
|
||||
|
||||
```sh
|
||||
eynollah binarization \
|
||||
-i <single image file> | -di <directory containing image files> \
|
||||
-o <output directory> \
|
||||
-m <directory containing model files>
|
||||
```
|
||||
* For some documents, while the quality is good, their scale is very large, and the performance of tool decreases. In such cases you can set `-as` (**a**llow **s**caling) to `true`. With this option enabled, the tool will try to rescale the image and only then the layout detection process will begin.
|
||||
|
||||
### Image Enhancement
|
||||
TODO
|
||||
* If you care about drop capitals (initials) and headings, you can set `-fl` (**f**ull **l**ayout) to `true`. With this setting, the tool can currently distinguish 7 document layout classes/elements.
|
||||
|
||||
### OCR
|
||||
* In cases where the document includes curved headers or curved lines, rectangular bounding boxes for textlines will not be a great option. In such cases it is strongly recommended setting the flag `-cl` (**c**urved **l**ines) to `true` to find contours of curved lines instead of rectangular bounding boxes. Be advised that enabling this option increases the processing time of the tool.
|
||||
|
||||
The OCR module performs text recognition using either a CNN-RNN model or a Transformer model.
|
||||
* To crop and save image regions inside the document, set the parameter `-si` (**s**ave **i**mages) to true and provide a directory path to store the extracted images.
|
||||
|
||||
The command-line interface for OCR can be called like this:
|
||||
|
||||
```sh
|
||||
eynollah ocr \
|
||||
-i <single image file> | -di <directory containing image files> \
|
||||
-dx <directory of xmls> \
|
||||
-o <output directory> \
|
||||
-m <directory containing model files> | --model_name <path to specific model>
|
||||
```
|
||||
|
||||
The following options can be used to further configure the ocr processing:
|
||||
|
||||
| option | description |
|
||||
|-------------------|:-------------------------------------------------------------------------------------------|
|
||||
| `-dib` | directory of binarized images (file type must be '.png'), prediction with both RGB and bin |
|
||||
| `-doit` | directory for output images rendered with the predicted text |
|
||||
| `--model_name` | file path to use specific model for OCR |
|
||||
| `-trocr` | use transformer ocr model (otherwise cnn_rnn model is used) |
|
||||
| `-etit` | export textline images and text in xml to output dir (OCR training data) |
|
||||
| `-nmtc` | cropped textline images will not be masked with textline contour |
|
||||
| `-bs` | ocr inference batch size. Default batch size is 2 for trocr and 8 for cnn_rnn models |
|
||||
| `-ds_pref` | add an abbrevation of dataset name to generated training data |
|
||||
| `-min_conf` | minimum OCR confidence value. OCR with textline conf lower than this will be ignored |
|
||||
|
||||
|
||||
### Reading Order Detection
|
||||
Reading order detection can be performed either as part of layout analysis based on image input, or, currently under
|
||||
development, based on pre-existing layout analysis data in PAGE-XML format as input.
|
||||
|
||||
The reading order detection module employs a pretrained model to identify the reading order from layouts represented in PAGE-XML files.
|
||||
|
||||
The command-line interface for machine based reading order can be called like this:
|
||||
|
||||
```sh
|
||||
eynollah machine-based-reading-order \
|
||||
-i <single image file> | -di <directory containing image files> \
|
||||
-xml <xml file name> | -dx <directory containing xml files> \
|
||||
-m <path to directory containing model files> \
|
||||
-o <output directory>
|
||||
```
|
||||
|
||||
## Use as OCR-D processor
|
||||
|
||||
See [`ocrd.md`](https://github.com/qurator-spk/eynollah/tree/main/docs/ocrd.md).
|
||||
|
||||
## How to cite
|
||||
|
||||
```bibtex
|
||||
@inproceedings{hip23rezanezhad,
|
||||
title = {Document Layout Analysis with Deep Learning and Heuristics},
|
||||
author = {Rezanezhad, Vahid and Baierer, Konstantin and Gerber, Mike and Labusch, Kai and Neudecker, Clemens},
|
||||
booktitle = {Proceedings of the 7th International Workshop on Historical Document Imaging and Processing {HIP} 2023,
|
||||
San José, CA, USA, August 25-26, 2023},
|
||||
publisher = {Association for Computing Machinery},
|
||||
address = {New York, NY, USA},
|
||||
year = {2023},
|
||||
pages = {73--78},
|
||||
url = {https://doi.org/10.1145/3604951.3605513}
|
||||
}
|
||||
```
|
||||
* This tool is actively being developed. If problems occur, or the performance does not meet your expectations, we welcome your feedback via [issues](https://github.com/qurator-spk/eynollah/issues).
|
||||
|
|
|
|||
|
|
@ -1,43 +0,0 @@
|
|||
## Inference with Docker
|
||||
|
||||
docker pull ghcr.io/qurator-spk/eynollah:latest
|
||||
|
||||
### 1. ocrd resource manager
|
||||
(just once, to get the models and install them into a named volume for later re-use)
|
||||
|
||||
vol_models=ocrd-resources:/usr/local/share/ocrd-resources
|
||||
docker run --rm -v $vol_models ocrd/eynollah ocrd resmgr download ocrd-eynollah-segment default
|
||||
|
||||
Now, each time you want to use Eynollah, pass the same resources volume again.
|
||||
Also, bind-mount some data directory, e.g. current working directory $PWD (/data is default working directory in the container).
|
||||
|
||||
Either use standalone CLI (2) or OCR-D CLI (3):
|
||||
|
||||
### 2. standalone CLI
|
||||
(follow self-help, cf. readme)
|
||||
|
||||
docker run --rm -v $vol_models -v $PWD:/data ocrd/eynollah eynollah binarization --help
|
||||
docker run --rm -v $vol_models -v $PWD:/data ocrd/eynollah eynollah layout --help
|
||||
docker run --rm -v $vol_models -v $PWD:/data ocrd/eynollah eynollah ocr --help
|
||||
|
||||
### 3. OCR-D CLI
|
||||
(follow self-help, cf. readme and https://ocr-d.de/en/spec/cli)
|
||||
|
||||
docker run --rm -v $vol_models -v $PWD:/data ocrd/eynollah ocrd-eynollah-segment -h
|
||||
docker run --rm -v $vol_models -v $PWD:/data ocrd/eynollah ocrd-sbb-binarize -h
|
||||
|
||||
Alternatively, just "log in" to the container once and use the commands there:
|
||||
|
||||
docker run --rm -v $vol_models -v $PWD:/data -it ocrd/eynollah bash
|
||||
|
||||
## Training with Docker
|
||||
|
||||
Build the Docker training image
|
||||
|
||||
cd train
|
||||
docker build -t model-training .
|
||||
|
||||
Run the Docker training image
|
||||
|
||||
cd train
|
||||
docker run --gpus all -v $PWD:/entry_point_dir model-training
|
||||
|
|
@ -1,18 +0,0 @@
|
|||
# Examples
|
||||
|
||||
Example outputs of various Eynollah models
|
||||
|
||||
# Binarisation
|
||||
|
||||
<img src="https://user-images.githubusercontent.com/952378/63592437-e433e400-c5b1-11e9-9c2d-889c6e93d748.jpg" width="45%"><img src="https://user-images.githubusercontent.com/952378/63592435-e433e400-c5b1-11e9-88e4-3e441b61fa67.jpg" width="45%">
|
||||
<img src="https://user-images.githubusercontent.com/952378/63592440-e4cc7a80-c5b1-11e9-8964-2cd1b22c87be.jpg" width="45%"><img src="https://user-images.githubusercontent.com/952378/63592438-e4cc7a80-c5b1-11e9-86dc-a9e9f8555422.jpg" width="45%">
|
||||
|
||||
# Reading Order Detection
|
||||
|
||||
<img src="https://github.com/user-attachments/assets/42df2582-4579-415e-92f1-54858a02c830" alt="Input Image" width="45%">
|
||||
<img src="https://github.com/user-attachments/assets/77fc819e-6302-4fc9-967c-ee11d10d863e" alt="Output Image" width="45%">
|
||||
|
||||
# OCR
|
||||
|
||||
<img src="https://github.com/user-attachments/assets/71054636-51c6-4117-b3cf-361c5cda3528" alt="Input Image" width="45%"><img src="https://github.com/user-attachments/assets/cfb3ce38-007a-4037-b547-21324a7d56dd" alt="Output Image" width="45%">
|
||||
<img src="https://github.com/user-attachments/assets/343b2ed8-d818-4d4a-b301-f304cbbebfcd" alt="Input Image" width="45%"><img src="https://github.com/user-attachments/assets/accb5ba7-e37f-477e-84aa-92eafa0d136e" alt="Output Image" width="45%">
|
||||
226
docs/models.md
226
docs/models.md
|
|
@ -1,226 +0,0 @@
|
|||
# Models documentation
|
||||
|
||||
This suite of 15 models presents a document layout analysis (DLA) system for historical documents implemented by
|
||||
pixel-wise segmentation using a combination of a ResNet50 encoder with various U-Net decoders. In addition, heuristic
|
||||
methods are applied to detect marginals and to determine the reading order of text regions.
|
||||
|
||||
The detection and classification of multiple classes of layout elements such as headings, images, tables etc. as part of
|
||||
DLA is required in order to extract and process them in subsequent steps. Altogether, the combination of image
|
||||
detection, classification and segmentation on the wide variety that can be found in over 400 years of printed cultural
|
||||
heritage makes this a very challenging task. Deep learning models are complemented with heuristics for the detection of
|
||||
text lines, marginals, and reading order. Furthermore, an optional image enhancement step was added in case of documents
|
||||
that either have insufficient pixel density and/or require scaling. Also, a column classifier for the analysis of
|
||||
multi-column documents was added. With these additions, DLA performance was improved, and a high accuracy in the
|
||||
prediction of the reading order is accomplished.
|
||||
|
||||
Two Arabic/Persian terms form the name of the model suite: عين الله, which can be transcribed as "ain'allah" or
|
||||
"eynollah"; it translates into English as "God's Eye" -- it sees (nearly) everything on the document image.
|
||||
|
||||
See the flowchart below for the different stages and how they interact:
|
||||
|
||||
<img width="810" height="691" alt="eynollah_flowchart" src="https://github.com/user-attachments/assets/42dd55bc-7b85-4b46-9afe-15ff712607f0" />
|
||||
|
||||
|
||||
|
||||
## Models
|
||||
|
||||
### Image enhancement
|
||||
|
||||
Model card: [Image Enhancement](https://huggingface.co/SBB/eynollah-enhancement)
|
||||
|
||||
This model addresses image resolution, specifically targeting documents with suboptimal resolution. In instances where
|
||||
the detection of document layout exhibits inadequate performance, the proposed enhancement aims to significantly improve
|
||||
the quality and clarity of the images, thus facilitating enhanced visual interpretation and analysis.
|
||||
|
||||
### Page extraction / border detection
|
||||
|
||||
Model card: [Page Extraction/Border Detection](https://huggingface.co/SBB/eynollah-page-extraction)
|
||||
|
||||
A problem that can negatively affect OCR are black margins around a page caused by document scanning. A deep learning
|
||||
model helps to crop to the page borders by using a pixel-wise segmentation method.
|
||||
|
||||
### Column classification
|
||||
|
||||
Model card: [Column Classification](https://huggingface.co/SBB/eynollah-column-classifier)
|
||||
|
||||
This model is a trained classifier that recognizes the number of columns in a document by use of a training set with
|
||||
manual classification of all documents into six classes with either one, two, three, four, five, or six and more columns
|
||||
respectively.
|
||||
|
||||
### Binarization
|
||||
|
||||
Model card: [Binarization](https://huggingface.co/SBB/eynollah-binarization)
|
||||
|
||||
This model is designed to tackle the intricate task of document image binarization, which involves segmentation of the
|
||||
image into white and black pixels. This process significantly contributes to the overall performance of the layout
|
||||
models, particularly in scenarios where the documents are degraded or exhibit subpar quality. The robust binarization
|
||||
capability of the model enables improved accuracy and reliability in subsequent layout analysis, thereby facilitating
|
||||
enhanced document understanding and interpretation.
|
||||
|
||||
### Main region detection
|
||||
|
||||
Model card: [Main Region Detection](https://huggingface.co/SBB/eynollah-main-regions)
|
||||
|
||||
This model has employed a different set of labels, including an artificial class specifically designed to encompass the
|
||||
text regions. The inclusion of this artificial class facilitates easier isolation of text regions by the model. This
|
||||
approach grants the advantage of training the model using downscaled images, which in turn leads to faster predictions
|
||||
during the inference phase. By incorporating this methodology, improved efficiency is achieved without compromising the
|
||||
model's ability to accurately identify and classify text regions within documents.
|
||||
|
||||
### Main region detection (with scaling augmentation)
|
||||
|
||||
Model card: [Main Region Detection (with scaling augmentation)](https://huggingface.co/SBB/eynollah-main-regions-aug-scaling)
|
||||
|
||||
Utilizing scaling augmentation, this model leverages the capability to effectively segment elements of extremely high or
|
||||
low scales within documents. By harnessing this technique, the tool gains a significant advantage in accurately
|
||||
categorizing and isolating such elements, thereby enhancing its overall performance and enabling precise analysis of
|
||||
documents with varying scale characteristics.
|
||||
|
||||
### Main region detection (with rotation augmentation)
|
||||
|
||||
Model card: [Main Region Detection (with rotation augmentation)](https://huggingface.co/SBB/eynollah-main-regions-aug-rotation)
|
||||
|
||||
This model takes advantage of rotation augmentation. This helps the tool to segment the vertical text regions in a
|
||||
robust way.
|
||||
|
||||
### Main region detection (ensembled)
|
||||
|
||||
Model card: [Main Region Detection (ensembled)](https://huggingface.co/SBB/eynollah-main-regions-ensembled)
|
||||
|
||||
The robustness of this model is attained through an ensembling technique that combines the weights from various epochs.
|
||||
By employing this approach, the model achieves a high level of resilience and stability, effectively leveraging the
|
||||
strengths of multiple epochs to enhance its overall performance and deliver consistent and reliable results.
|
||||
|
||||
### Full region detection (1,2-column documents)
|
||||
|
||||
Model card: [Full Region Detection (1,2-column documents)](https://huggingface.co/SBB/eynollah-full-regions-1column)
|
||||
|
||||
This model deals with documents comprising of one and two columns.
|
||||
|
||||
### Full region detection (3,n-column documents)
|
||||
|
||||
Model card: [Full Region Detection (3,n-column documents)](https://huggingface.co/SBB/eynollah-full-regions-3pluscolumn)
|
||||
|
||||
This model is responsible for detecting headers and drop capitals in documents with three or more columns.
|
||||
|
||||
### Textline detection
|
||||
|
||||
Model card: [Textline Detection](https://huggingface.co/SBB/eynollah-textline)
|
||||
|
||||
The method for textline detection combines deep learning and heuristics. In the deep learning part, an image-to-image
|
||||
model performs binary segmentation of the document into the classes textline vs. background. In the heuristics part,
|
||||
bounding boxes or contours are derived from binary segmentation.
|
||||
|
||||
Skewed documents can heavily affect textline detection accuracy, so robust deskewing is needed. But detecting textlines
|
||||
with rectangle bounding boxes cannot deal with partially curved textlines. To address this, a functionality
|
||||
specifically for documents with curved textlines was included. After finding the contour of a text region and its
|
||||
corresponding textline segmentation, the text region is cut into smaller vertical straps. For each strap, its textline
|
||||
segmentation is first deskewed and then the textlines are separated with the same heuristic method as for finding
|
||||
textline bounding boxes. Later, the strap is rotated back into its original orientation.
|
||||
|
||||
### Textline detection (light)
|
||||
|
||||
Model card: [Textline Detection Light (simpler but faster method)](https://huggingface.co/SBB/eynollah-textline_light)
|
||||
|
||||
The method for textline detection combines deep learning and heuristics. In the deep learning part, an image-to-image
|
||||
model performs binary segmentation of the document into the classes textline vs. background. In the heuristics part,
|
||||
bounding boxes or contours are derived from binary segmentation.
|
||||
|
||||
In the context of this textline model, a distinct labeling approach has been employed to ensure accurate predictions.
|
||||
Specifically, an artificial bounding class has been incorporated alongside the textline classes. This strategic
|
||||
inclusion effectively prevents any spurious connections between adjacent textlines during the prediction phase, thereby
|
||||
enhancing the model's ability to accurately identify and delineate individual textlines within documents. This model
|
||||
eliminates the need for additional heuristics in extracting textline contours.
|
||||
|
||||
### Table detection
|
||||
|
||||
Model card: [Table Detection](https://huggingface.co/SBB/eynollah-tables)
|
||||
|
||||
The objective of this model is to perform table segmentation in historical document images. Due to the pixel-wise
|
||||
segmentation approach employed and the presence of traditional tables predominantly composed of text, the detection of
|
||||
tables required the incorporation of heuristics to achieve reasonable performance. These heuristics were necessary to
|
||||
effectively identify and delineate tables within the historical document images, ensuring accurate segmentation and
|
||||
enabling subsequent analysis and interpretation.
|
||||
|
||||
### Image detection
|
||||
|
||||
Model card: [Image Detection](https://huggingface.co/SBB/eynollah-image-extraction)
|
||||
|
||||
This model is used for the task of illustration detection only.
|
||||
|
||||
### Reading order detection
|
||||
|
||||
Model card: [Reading Order Detection]()
|
||||
|
||||
The model extracts the reading order of text regions from the layout by classifying pairwise relationships between them. A sorting algorithm then determines the overall reading sequence.
|
||||
|
||||
### OCR
|
||||
|
||||
We have trained three OCR models: two CNN-RNN–based models and one transformer-based TrOCR model. The CNN-RNN models are generally faster and provide better results in most cases, though their performance decreases with heavily degraded images. The TrOCR model, on the other hand, is computationally expensive and slower during inference, but it can possibly produce better results on strongly degraded images.
|
||||
|
||||
#### CNN-RNN model: model_eynollah_ocr_cnnrnn_20250805
|
||||
|
||||
This model is trained on data where most of the samples are in Fraktur german script.
|
||||
|
||||
| Dataset | Input | CER | WER |
|
||||
|-----------------------|:-------|:-----------|:----------|
|
||||
| OCR-D-GT-Archiveform | BIN | 0.02147 | 0.05685 |
|
||||
| OCR-D-GT-Archiveform | RGB | 0.01636 | 0.06285 |
|
||||
|
||||
#### CNN-RNN model: model_eynollah_ocr_cnnrnn_20250904 (Default)
|
||||
|
||||
Compared to the model_eynollah_ocr_cnnrnn_20250805 model, this model is trained on a larger proportion of Antiqua data and achieves superior performance.
|
||||
|
||||
| Dataset | Input | CER | WER |
|
||||
|-----------------------|:------------|:-----------|:----------|
|
||||
| OCR-D-GT-Archiveform | BIN | 0.01635 | 0.05410 |
|
||||
| OCR-D-GT-Archiveform | RGB | 0.01471 | 0.05813 |
|
||||
| BLN600 | RGB | 0.04409 | 0.08879 |
|
||||
| BLN600 | Enhanced | 0.03599 | 0.06244 |
|
||||
|
||||
|
||||
#### Transformer OCR model: model_eynollah_ocr_trocr_20250919
|
||||
|
||||
This transformer OCR model is trained on the same data as model_eynollah_ocr_trocr_20250919.
|
||||
|
||||
| Dataset | Input | CER | WER |
|
||||
|-----------------------|:------------|:-----------|:----------|
|
||||
| OCR-D-GT-Archiveform | BIN | 0.01841 | 0.05589 |
|
||||
| OCR-D-GT-Archiveform | RGB | 0.01552 | 0.06177 |
|
||||
| BLN600 | RGB | 0.06347 | 0.13853 |
|
||||
|
||||
##### Qualitative evaluation of the models
|
||||
|
||||
| <img width="1600" src="https://github.com/user-attachments/assets/120fec0c-c370-46a6-b132-b0af800607cf"> | <img width="1000" src="https://github.com/user-attachments/assets/d84e6819-0a2a-4b3a-bb7d-ceac941babc4"> | <img width="1000" src="https://github.com/user-attachments/assets/bdd27cdb-bbec-4223-9a86-de7a27c6d018"> | <img width="1000" src="https://github.com/user-attachments/assets/1a507c75-75de-4da3-9545-af3746b9a207"> |
|
||||
|:---:|:---:|:---:|:---:|
|
||||
| Image | cnnrnn_20250805 | cnnrnn_20250904 | trocr_20250919 |
|
||||
|
||||
|
||||
|
||||
| <img width="2000" src="https://github.com/user-attachments/assets/9bc13d48-2a92-45fc-88db-c07ffadba067"> | <img width="1000" src="https://github.com/user-attachments/assets/2b294aeb-1362-4d6e-b70f-8aeffd94c5e7"> | <img width="1000" src="https://github.com/user-attachments/assets/9911317e-632e-4e6a-8839-1fb7e783da11"> | <img width="1000" src="https://github.com/user-attachments/assets/2c5626d9-0d23-49d3-80f5-a95f629c9c76"> |
|
||||
|:---:|:---:|:---:|:---:|
|
||||
| Image | cnnrnn_20250805 | cnnrnn_20250904 | trocr_20250919 |
|
||||
|
||||
|
||||
| <img width="2000" src="https://github.com/user-attachments/assets/d54d8510-5c6a-4ab0-9ba7-f6ec4ad452c6"> | <img width="1000" src="https://github.com/user-attachments/assets/a418b25b-00dc-493a-b3a3-b325b9b0cb85"> | <img width="1000" src="https://github.com/user-attachments/assets/df6e2b9e-a821-4b4c-8868-0c765700c341"> | <img width="1000" src="https://github.com/user-attachments/assets/b90277f5-40f4-4c99-80a2-da400f7d3640"> |
|
||||
|:---:|:---:|:---:|:---:|
|
||||
| Image | cnnrnn_20250805 | cnnrnn_20250904 | trocr_20250919 |
|
||||
|
||||
|
||||
| <img width="2000" src="https://github.com/user-attachments/assets/7ec49211-099f-4c21-9e60-47bfdf21f1b6"> | <img width="1000" src="https://github.com/user-attachments/assets/00ef9785-8885-41b3-bf6e-21eab743df71"> | <img width="1000" src="https://github.com/user-attachments/assets/13eb9f62-4d5a-46dc-befc-b02eb4f31fc1"> | <img width="1000" src="https://github.com/user-attachments/assets/a5c078d1-6d15-4d12-9040-526d7063d459"> |
|
||||
|:---:|:---:|:---:|:---:|
|
||||
| Image | cnnrnn_20250805 | cnnrnn_20250904 | trocr_20250919 |
|
||||
|
||||
|
||||
|
||||
## Heuristic methods
|
||||
|
||||
Additionally, some heuristic methods are employed to further improve the model predictions:
|
||||
|
||||
* After border detection, the largest contour is determined by a bounding box, and the image cropped to these coordinates.
|
||||
* Unlike the non-light version, where the image is scaled up to help the model better detect the background spaces between text regions, the light version uses down-scaled images. In this case, introducing an artificial class along the boundaries of text regions and text lines has helped to isolate and separate the text regions more effectively.
|
||||
* A minimum area is defined for text regions in relation to the overall image dimensions, so that very small regions that are noise can be filtered out.
|
||||
* In the non-light version, deskewing is applied at the text-region level (since regions may have different degrees of skew) to improve text-line segmentation results. In contrast, the light version performs deskewing only at the page level to enhance margin detection and heuristic reading-order estimation.
|
||||
* After deskewing, a calculation of the pixel distribution on the X-axis allows the separation of textlines (foreground) and background pixels (only in non-light version).
|
||||
* Finally, using the derived coordinates, bounding boxes are determined for each textline (only in non-light version).
|
||||
* As mentioned above, the reading order can be determined using a model; however, this approach is computationally expensive, time-consuming, and less accurate due to the limited amount of ground-truth data available for training. Therefore, our tool uses a heuristic reading-order detection method as the default. The heuristic approach relies on headers and separators to determine the reading order of text regions.
|
||||
26
docs/ocrd.md
26
docs/ocrd.md
|
|
@ -1,26 +0,0 @@
|
|||
## Use as OCR-D processor
|
||||
|
||||
Eynollah ships with a CLI interface to be used as [OCR-D](https://ocr-d.de) [processor](https://ocr-d.de/en/spec/cli),
|
||||
formally described in [`ocrd-tool.json`](https://github.com/qurator-spk/eynollah/tree/main/src/eynollah/ocrd-tool.json).
|
||||
|
||||
When using Eynollah in OCR-D, the source image file group with (preferably) RGB images should be used as input like this:
|
||||
|
||||
ocrd-eynollah-segment -I OCR-D-IMG -O OCR-D-SEG -P models eynollah_layout_v0_5_0
|
||||
|
||||
If the input file group is PAGE-XML (from a previous OCR-D workflow step), Eynollah behaves as follows:
|
||||
- existing regions are kept and ignored (i.e. in effect they might overlap segments from Eynollah results)
|
||||
- existing annotation (and respective `AlternativeImage`s) are partially _ignored_:
|
||||
- previous page frame detection (`cropped` images)
|
||||
- previous derotation (`deskewed` images)
|
||||
- previous thresholding (`binarized` images)
|
||||
- if the page-level image nevertheless deviates from the original (`@imageFilename`)
|
||||
(because some other preprocessing step was in effect like `denoised`), then
|
||||
the output PAGE-XML will be based on that as new top-level (`@imageFilename`)
|
||||
|
||||
ocrd-eynollah-segment -I OCR-D-XYZ -O OCR-D-SEG -P models eynollah_layout_v0_5_0
|
||||
|
||||
In general, it makes more sense to add other workflow steps **after** Eynollah.
|
||||
|
||||
There is also an OCR-D processor for binarization:
|
||||
|
||||
ocrd-sbb-binarize -I OCR-D-IMG -O OCR-D-BIN -P models default-2021-03-09
|
||||
804
docs/train.md
804
docs/train.md
|
|
@ -1,804 +0,0 @@
|
|||
# Prerequisistes
|
||||
|
||||
## 1. Install Eynollah with training dependencies
|
||||
|
||||
Clone the repository and install eynollah along with the dependencies necessary for training:
|
||||
|
||||
```sh
|
||||
git clone https://github.com/qurator-spk/eynollah
|
||||
cd eynollah
|
||||
pip install '.[training]'
|
||||
```
|
||||
|
||||
## 2. Pretrained encoder
|
||||
|
||||
Download our pretrained weights and add them to a `train/pretrained_model` folder:
|
||||
|
||||
```sh
|
||||
cd train
|
||||
wget -O pretrained_model.tar.gz https://zenodo.org/records/17243320/files/pretrained_model_v0_5_1.tar.gz?download=1
|
||||
tar xf pretrained_model.tar.gz
|
||||
```
|
||||
|
||||
## 3. Example data
|
||||
|
||||
### Binarization
|
||||
A small sample of training data for binarization experiment can be found on [Zenodo](https://zenodo.org/records/17243320/files/training_data_sample_binarization_v0_5_1.tar.gz?download=1),
|
||||
which contains `images` and `labels` folders.
|
||||
|
||||
## 4. Helpful tools
|
||||
|
||||
* [`pagexml2img`](https://github.com/qurator-spk/page2img)
|
||||
> Tool to extract 2-D or 3-D RGB images from PAGE-XML data. In the former case, the output will be 1 2-D image array which each class has filled with a pixel value. In the case of a 3-D RGB image,
|
||||
each class will be defined with a RGB value and beside images, a text file of classes will also be produced.
|
||||
* [`cocoSegmentationToPng`](https://github.com/nightrome/cocostuffapi/blob/17acf33aef3c6cc2d6aca46dcf084266c2778cf0/PythonAPI/pycocotools/cocostuffhelper.py#L130)
|
||||
> Convert COCO GT or results for a single image to a segmentation map and write it to disk.
|
||||
* [`ocrd-segment-extract-pages`](https://github.com/OCR-D/ocrd_segment/blob/master/ocrd_segment/extract_pages.py)
|
||||
> Extract region classes and their colours in mask (pseg) images. Allows the color map as free dict parameter, and comes with a default that mimics PageViewer's coloring for quick debugging; it also warns when regions do overlap.
|
||||
|
||||
# Training documentation
|
||||
|
||||
This document aims to assist users in preparing training datasets, training models, and
|
||||
performing inference with trained models. We cover various use cases including
|
||||
pixel-wise segmentation, image classification, image enhancement, and
|
||||
machine-based reading order detection. For each use case, we provide guidance
|
||||
on how to generate the corresponding training dataset.
|
||||
|
||||
The following three tasks can all be accomplished using the code in the
|
||||
[`train`](https://github.com/qurator-spk/eynollah/tree/main/train) directory:
|
||||
|
||||
* [Generate training dataset](#generate-training-dataset)
|
||||
* [Train a model](#train-a-model)
|
||||
* [Inference with the trained model](#inference-with-the-trained-model)
|
||||
|
||||
## Training, evaluation and output
|
||||
|
||||
The train and evaluation folders should contain subfolders of `images` and `labels`.
|
||||
|
||||
The output folder should be an empty folder where the output model will be written to.
|
||||
|
||||
## Generate training dataset
|
||||
|
||||
The script `generate_gt_for_training.py` is used for generating training datasets. As the results of the following
|
||||
command demonstrates, the dataset generator provides several subcommands:
|
||||
|
||||
```sh
|
||||
eynollah-training generate-gt --help
|
||||
```
|
||||
|
||||
The three most important subcommands are:
|
||||
|
||||
* image-enhancement
|
||||
* machine-based-reading-order
|
||||
* pagexml2label
|
||||
|
||||
### image-enhancement
|
||||
|
||||
Generating a training dataset for image enhancement is quite straightforward. All that is needed is a set of
|
||||
high-resolution images. The training dataset can then be generated using the following command:
|
||||
|
||||
```sh
|
||||
eynollah-training image-enhancement \
|
||||
-dis "dir of high resolution images" \
|
||||
-dois "dir where degraded images will be written" \
|
||||
-dols "dir where the corresponding high resolution image will be written as label" \
|
||||
-scs "degrading scales json file"
|
||||
```
|
||||
|
||||
The scales JSON file is a dictionary with a key named `scales` and values representing scales smaller than 1. Images are
|
||||
downscaled based on these scales and then upscaled again to their original size. This process causes the images to lose
|
||||
resolution at different scales. The degraded images are used as input images, and the original high-resolution images
|
||||
serve as labels. The enhancement model can be trained with this generated dataset. The scales JSON file looks like this:
|
||||
|
||||
```yaml
|
||||
{
|
||||
"scales": [0.5, 0.55, 0.6, 0.65, 0.7, 0.75, 0.8, 0.85, 0.9]
|
||||
}
|
||||
```
|
||||
|
||||
### machine-based-reading-order
|
||||
|
||||
For machine-based reading order, we aim to determine the reading priority between two sets of text regions. The model's
|
||||
input is a three-channel image: the first and last channels contain information about each of the two text regions,
|
||||
while the middle channel encodes prominent layout elements necessary for reading order, such as separators and headers.
|
||||
To generate the training dataset, our script requires a PAGE XML file that specifies the image layout with the correct
|
||||
reading order.
|
||||
|
||||
For output images, it is necessary to specify the width and height. Additionally, a minimum text region size can be set
|
||||
to filter out regions smaller than this minimum size. This minimum size is defined as the ratio of the text region area
|
||||
to the image area, with a default value of zero. To run the dataset generator, use the following command:
|
||||
|
||||
```shell
|
||||
eynollah-training generate-gt machine-based-reading-order \
|
||||
-dx "dir of GT xml files" \
|
||||
-domi "dir where output images will be written" \
|
||||
"" -docl "dir where the labels will be written" \
|
||||
-ih "height" \
|
||||
-iw "width" \
|
||||
-min "min area ratio"
|
||||
```
|
||||
|
||||
### pagexml2label
|
||||
|
||||
`pagexml2label` is designed to generate labels from PAGE XML GT files for various pixel-wise segmentation use cases,
|
||||
including:
|
||||
- `printspace` (i.e. page frame),
|
||||
- `layout` (i.e. regions),
|
||||
- `textline`,
|
||||
- `word`, and
|
||||
- `glyph`.
|
||||
|
||||
To train a pixel-wise segmentation model, we require images along with their corresponding labels. Our training script
|
||||
expects a PNG image where each pixel corresponds to a label, represented by an integer. The background is always labeled
|
||||
as zero, while other elements are assigned different integers. For instance, if we have ground truth data with four
|
||||
elements including the background, the classes would be labeled as 0, 1, 2, and 3 respectively.
|
||||
|
||||
In binary segmentation scenarios such as textline or page extraction, the background is encoded as 0, and the desired
|
||||
element is automatically encoded as 1 in the PNG label.
|
||||
|
||||
To specify the desired use case and the elements to be extracted in the PNG labels, a custom JSON file can be passed.
|
||||
For example, in the case of textline detection, the JSON contents could be this:
|
||||
|
||||
```yaml
|
||||
{
|
||||
"use_case": "textline"
|
||||
}
|
||||
```
|
||||
|
||||
In the case of layout segmentation, the config JSON file might look like this:
|
||||
|
||||
```yaml
|
||||
{
|
||||
"use_case": "layout",
|
||||
"textregions": {"rest_as_paragraph": 1, "drop-capital": 1, "header": 2, "heading": 2, "marginalia": 3},
|
||||
"imageregion": 4,
|
||||
"separatorregion": 5,
|
||||
"graphicregions": {"rest_as_decoration": 6, "stamp": 7}
|
||||
}
|
||||
```
|
||||
|
||||
The same example if `PrintSpace` (or `Border`) should be represented as a unique class:
|
||||
|
||||
```yaml
|
||||
{
|
||||
"use_case": "layout",
|
||||
"textregions": {"rest_as_paragraph": 1, "drop-capital": 1, "header": 2, "heading": 2, "marginalia": 3},
|
||||
"imageregion": 4,
|
||||
"separatorregion": 5,
|
||||
"graphicregions": {"rest_as_decoration": 6, "stamp": 7}
|
||||
"printspace_as_class_in_layout": 8
|
||||
}
|
||||
```
|
||||
|
||||
In the `layout` use-case, it is beneficial to first understand the structure of the PAGE XML file and its elements.
|
||||
For a given page image, the visible segments are annotated in XML with their polygon coordinates and types.
|
||||
On the region level, available segment types include `TextRegion`, `SeparatorRegion`, `ImageRegion`, `GraphicRegion`,
|
||||
`NoiseRegion` and `TableRegion`.
|
||||
|
||||
Moreover, text regions and graphic regions in particular are subdivided via `@type`:
|
||||
- The allowed subtypes for text regions are `paragraph`, `heading`, `marginalia`, `drop-capital`, `header`, `footnote`,
|
||||
`footnote-continued`, `signature-mark`, `page-number` and `catch-word`.
|
||||
- The known subtypes for graphic regions are `handwritten-annotation`, `decoration`, `stamp` and `signature`.
|
||||
|
||||
These types and subtypes must be mapped to classes for the segmentation model. However, sometimes these fine-grained
|
||||
distinctions are not useful or the existing annotations are not very usable (too scarce or too unreliable).
|
||||
In that case, instead of these subtypes with a specific mapping, they can be pooled together by using the two special
|
||||
types:
|
||||
- `rest_as_paragraph` (mapping missing TextRegion subtypes and `paragraph`)
|
||||
- `rest_as_decoration` (mapping missing GraphicRegion subtypes and `decoration`)
|
||||
|
||||
(That way, users can extract all known types from the labels and be confident that no subtypes are overlooked.)
|
||||
|
||||
In the custom JSON example shown above, `header` and `heading` are extracted as the same class,
|
||||
while `marginalia` is modelled as a different class. All other text region types, including `drop-capital`,
|
||||
are grouped into the same class. For graphic regions, `stamp` has its own class, while all other types
|
||||
are classified together. `ImageRegion` and `SeparatorRegion` will also represented with a class label in the
|
||||
training data. However, other regions like `NoiseRegion` or `TableRegion` will not be included in the PNG files,
|
||||
even if they were present in the PAGE XML.
|
||||
|
||||
The tool expects various command-line options:
|
||||
|
||||
```sh
|
||||
eynollah-training generate-gt pagexml2label \
|
||||
-dx "dir of input PAGE XML files" \
|
||||
-do "dir of output label PNG files" \
|
||||
-cfg "custom config JSON file" \
|
||||
-to "output type (2d or 3d)"
|
||||
```
|
||||
|
||||
As output type, use
|
||||
- `2d` for training,
|
||||
- `3d` to just visualise the labels.
|
||||
|
||||
We have also defined an artificial class that can be added to (rendered around) the boundary
|
||||
of text region types or text lines in order to make separation of neighbouring segments more
|
||||
reliable. The key is called `artificial_class_on_boundary`, and it takes a list of text region
|
||||
types to be applied to.
|
||||
|
||||
Our example JSON config file could then look like this:
|
||||
|
||||
```yaml
|
||||
{
|
||||
"use_case": "layout",
|
||||
"textregions": {
|
||||
"paragraph": 1,
|
||||
"drop-capital": 1,
|
||||
"header": 2,
|
||||
"heading": 2,
|
||||
"marginalia": 3
|
||||
},
|
||||
"imageregion": 4,
|
||||
"separatorregion": 5,
|
||||
"graphicregions": {
|
||||
"rest_as_decoration": 6
|
||||
},
|
||||
"artificial_class_on_boundary": ["paragraph", "header", "heading", "marginalia"],
|
||||
"artificial_class_label": 7
|
||||
}
|
||||
```
|
||||
|
||||
This implies that the artificial class label (denoted by 7) will be present in the generated PNG files
|
||||
and will only be added around segments labeled `paragraph`, `header`, `heading` or `marginalia`. (This
|
||||
class will be handled specially during decoding at inference, and not show up in final results.)
|
||||
|
||||
For `printspace`, `textline`, `word`, and `glyph` segmentation use-cases, there is no `artificial_class_on_boundary` key,
|
||||
but `artificial_class_label` is available. If specified in the config file, then its value should be set at 2, because
|
||||
these elements represent binary classification problems (with background represented as 0, and segments as 1, respectively).
|
||||
|
||||
For example, the JSON config for textline detection could look as follows:
|
||||
|
||||
```yaml
|
||||
{
|
||||
"use_case": "textline",
|
||||
"artificial_class_label": 2
|
||||
}
|
||||
```
|
||||
|
||||
If the coordinates of `PrintSpace` (or `Border`) are present in the PAGE XML ground truth files,
|
||||
and one wishes to crop images to only cover the print space bounding box, this can be achieved
|
||||
by passing the `-ps` option. Note that in this scenario, the directory of the original images
|
||||
must also be provided, to ensure that the images are cropped in sync with the labels. The command
|
||||
line would then resemble this:
|
||||
|
||||
```sh
|
||||
eynollah-training generate-gt pagexml2label \
|
||||
-dx "dir of input PAGE XML files" \
|
||||
-do "dir of output label PNG files" \
|
||||
-cfg "custom config JSON file" \
|
||||
-to "output type (2d or 3d)" \
|
||||
-ps \
|
||||
-di "dir of input original images" \
|
||||
-doi "dir of output cropped images"
|
||||
```
|
||||
|
||||
Also, note that it can be detrimental to layout training if there are visible segments which
|
||||
the annotation does not account for (and thus the model must learn to ignore). So if the images
|
||||
are not cropped, the `-ps` _should_ be used. If a PAGE XML file is missing `PrintSpace` (or `Border`)
|
||||
annotations, use `-mps` to either `skip` these or `project` (i.e. crop from existing segments).
|
||||
|
||||
## Train a model
|
||||
|
||||
### classification
|
||||
|
||||
For the image classification use-case, we have not provided a ground truth generator, as it is unnecessary.
|
||||
All we require is a training directory with subdirectories, each containing images of its respective classes. We need
|
||||
separate directories for training and evaluation, and the class names (subdirectories) must be consistent across both
|
||||
directories. Additionally, the class names should be specified in the config JSON file, as shown in the following
|
||||
example. If, for instance, we aim to classify "apple" and "orange," with a total of 2 classes, the
|
||||
`classification_classes_name` key in the config file should appear as follows:
|
||||
|
||||
```yaml
|
||||
{
|
||||
"backbone_type" : "nontransformer",
|
||||
"task": "classification",
|
||||
"n_classes" : 2,
|
||||
"n_epochs" : 10,
|
||||
"input_height" : 448,
|
||||
"input_width" : 448,
|
||||
"weight_decay" : 1e-6,
|
||||
"n_batch" : 4,
|
||||
"learning_rate": 1e-4,
|
||||
"f1_threshold_classification": 0.8,
|
||||
"pretraining" : true,
|
||||
"classification_classes_name" : {"0":"apple", "1":"orange"},
|
||||
"dir_train": "./train",
|
||||
"dir_eval": "./eval",
|
||||
"dir_output": "./output"
|
||||
}
|
||||
```
|
||||
|
||||
Then `dir_train` should be like this:
|
||||
|
||||
```
|
||||
.
|
||||
└── train # train directory
|
||||
├── apple # directory of images for apple class
|
||||
└── orange # directory of images for orange class
|
||||
```
|
||||
|
||||
And `dir_eval` analogously:
|
||||
|
||||
```
|
||||
.
|
||||
└── eval # evaluation directory
|
||||
├── apple # directory of images for apple class
|
||||
└── orange # directory of images for orange class
|
||||
|
||||
```
|
||||
|
||||
The classification model can be trained using the following command line:
|
||||
|
||||
```sh
|
||||
eynollah-training train with config_classification.json
|
||||
```
|
||||
|
||||
As evident in the example JSON file above, for classification, we utilize a "f1_threshold_classification" parameter.
|
||||
This parameter is employed to gather all models with an evaluation f1 score surpassing this threshold. Subsequently,
|
||||
an ensemble of these model weights is executed, and a model is saved in the output directory as "model_ens_avg".
|
||||
Additionally, the weight of the best model based on the evaluation f1 score is saved as "model_best".
|
||||
|
||||
### reading order
|
||||
An example config json file for machine based reading order should be like this:
|
||||
|
||||
```yaml
|
||||
{
|
||||
"backbone_type" : "nontransformer",
|
||||
"task": "reading_order",
|
||||
"n_classes" : 1,
|
||||
"n_epochs" : 5,
|
||||
"input_height" : 672,
|
||||
"input_width" : 448,
|
||||
"weight_decay" : 1e-6,
|
||||
"n_batch" : 4,
|
||||
"learning_rate": 1e-4,
|
||||
"pretraining" : true,
|
||||
"dir_train": "./train",
|
||||
"dir_eval": "./eval",
|
||||
"dir_output": "./output"
|
||||
}
|
||||
```
|
||||
|
||||
The "dir_train" should be like this:
|
||||
|
||||
```
|
||||
.
|
||||
└── train # train directory
|
||||
├── images # directory of images
|
||||
└── labels # directory of labels
|
||||
```
|
||||
|
||||
And the "dir_eval" the same structure as train directory:
|
||||
|
||||
```
|
||||
.
|
||||
└── eval # evaluation directory
|
||||
├── images # directory of images
|
||||
└── labels # directory of labels
|
||||
```
|
||||
|
||||
The reading-order model can be trained like the classification case command line.
|
||||
|
||||
### Segmentation (Textline, Binarization, Page extraction and layout) and enhancement
|
||||
|
||||
#### Parameter configuration for segmentation or enhancement usecases
|
||||
|
||||
The following parameter configuration can be applied to all segmentation use cases and enhancements. The augmentation,
|
||||
its sub-parameters, and continued training are defined only for segmentation use cases and enhancements, not for
|
||||
classification and machine-based reading order, as you can see in their example config files.
|
||||
|
||||
* `task`: The task parameter must be one of the following values:
|
||||
- `binarization`,
|
||||
- `enhancement`,
|
||||
- `segmentation`,
|
||||
- `classification`,
|
||||
- `reading_order`.
|
||||
* `backbone_type`: For the tasks `segmentation` (such as text line, and region layout detection),
|
||||
`binarization` and `enhancement`, we offer two backbone options:
|
||||
- `nontransformer` (only a CNN ResNet-50).
|
||||
- `transformer` (first apply a CNN, followed by a transformer)
|
||||
* `transformer_cnn_first`: Whether to apply the CNN first (followed by the transformer) when using `transformer` backbone.
|
||||
* `transformer_num_patches_xy`: Number of patches for vision transformer in x and y direction respectively.
|
||||
* `transformer_patchsize_x`: Patch size of vision transformer patches in x direction.
|
||||
* `transformer_patchsize_y`: Patch size of vision transformer patches in y direction.
|
||||
* `transformer_projection_dim`: Transformer projection dimension. Default value is 64.
|
||||
* `transformer_mlp_head_units`: Transformer Multilayer Perceptron (MLP) head units. Default value is [128, 64].
|
||||
* `transformer_layers`: transformer layers. Default value is 8.
|
||||
* `transformer_num_heads`: Transformer number of heads. Default value is 4.
|
||||
* `patches`: Whether to break up (tile) input images into smaller patches (input size of the model).
|
||||
If `false`, the model will see the image once (resized to the input size of the model).
|
||||
Should be set to `false` for cases like page extraction.
|
||||
* `n_batch`: Number of batches at each iteration.
|
||||
* `n_classes`: Number of classes. In the case of binary classification this should be 2. In the case of reading_order it
|
||||
should set to 1. And for the case of layout detection just the unique number of classes should be given.
|
||||
* `n_epochs`: Number of epochs (iterations over the data) to train.
|
||||
* `input_height`: the image height for the model's input.
|
||||
* `input_width`: the image width for the model's input.
|
||||
* `weight_decay`: Weight decay of l2 regularization of model layers.
|
||||
* `weighted_loss`: If `true`, this means that you want to apply weighted categorical crossentropy as loss function.
|
||||
(Mutually exclusive with `is_loss_soft_dice`, and only applies for `segmentation` and `binarization` tasks.)
|
||||
* `pretraining`: Set to `true` to (download and) initialise pretrained weights of ResNet50 encoder.
|
||||
* `dir_train`: Path to directory of raw training data (as extracted via `pagexml2labels`, i.e. with subdirectories
|
||||
`images` and `labels` for input images and output labels.
|
||||
(These are not prepared for training the model, yet. Upon first run, the raw data will be transformed to suitable size
|
||||
needed for the model, and written in `dir_output` under `train` and `eval` subdirectories. See `data_is_provided`.)
|
||||
* `dir_eval`: Ditto for raw evaluation data.
|
||||
* `dir_output`: Directory to write model checkpoints, logs (for Tensorboard) and precomputed images to.
|
||||
* `data_is_provided`: If you have already trained at least one complete epoch (using the same data settings) before,
|
||||
you can set this to `true` to avoid computing the resized / patched / augmented image files again.
|
||||
Be sure that there are subdirectories `train` and `eval` data are in `dir_output` (each with subdirectories `images`
|
||||
and `labels`, respectively).
|
||||
* `continue_training`: If `true`, continue training a model checkpoint from a previous run.
|
||||
This requires providing the directory of the model checkpoint to load via `dir_of_start_model`
|
||||
and setting `index_start` counter for naming new checkpoints.
|
||||
For example if you have already trained for 3 epochs, then your last index is 2, so if you want
|
||||
to continue with `model_04`, `model_05` etc., set `index_start=3`.
|
||||
* `index_start`: Starting index for saving models in the case that `continue_training` is `true`.
|
||||
(Existing checkpoints above this will be overwritten.)
|
||||
* `dir_of_start_model`: Directory containing existing model checkpoint to initialise model weights from when `continue_training=true`.
|
||||
(Can be an epoch-interval checkpoint, or batch-interval checkpoint from `save_interval`.)
|
||||
* `augmentation`: If you want to apply any kind of augmentation this parameter should first set to `true`.
|
||||
The remaining settings pertain to that...
|
||||
* `flip_aug`: If `true`, different types of flipping over the image arrays. Requires `flip_index` parameter.
|
||||
* `flip_index`: List of flip codes (as in `cv2.flip`, i.e. 0 for vertical, positive for horizontal shift, negative for vertical and horizontal shift).
|
||||
* `blur_aug`: If `true`, different types of blurring will be applied on image. Requires `blur_k` parameter.
|
||||
* `blur_k`: Method of blurring (`gauss`, `median` or `blur`).
|
||||
* `scaling`: If `true`, scaling will be applied on image. Requires `scales` parameter.
|
||||
* `scales`: List of scale factors for scaling.
|
||||
* `scaling_bluring`: If `true`, combination of scaling and blurring will be applied on image.
|
||||
* `scaling_binarization`: If `true`, combination of scaling and binarization will be applied on image.
|
||||
* `scaling_flip`: If `true`, combination of scaling and flip will be applied on image.
|
||||
* `degrading`: If `true`, degrading will be applied to the image. Requires `degrade_scales` parameter.
|
||||
* `degrade_scales`: List of intensity factors for degrading.
|
||||
* `brightening`: If `true`, brightening will be applied to the image. Requires `brightness` parameter.
|
||||
* `brightness`: List of intensity factors for brightening.
|
||||
* `binarization`: If `true`, Otsu thresholding will be applied to augment the input data with binarized images.
|
||||
* `dir_img_bin`: With `binarization`, use this directory to read precomputed binarized images instead of ad-hoc Otsu.
|
||||
(Base names should correspond to the files in `dir_train/images`.)
|
||||
* `rotation`: If `true`, 90° rotation will be applied on images.
|
||||
* `rotation_not_90`: If `true`, random rotation (other than 90°) will be applied on image. Requires `thetha` parameter.
|
||||
* `thetha`: List of rotation angles (in degrees).
|
||||
|
||||
In case of segmentation and enhancement the train and evaluation data should be organised as follows.
|
||||
|
||||
The "dir_train" directory should be like this:
|
||||
|
||||
```
|
||||
.
|
||||
└── train # train directory
|
||||
├── images # directory of images
|
||||
└── labels # directory of labels
|
||||
```
|
||||
|
||||
And the "dir_eval" the same structure as train directory:
|
||||
|
||||
```
|
||||
.
|
||||
└── eval # evaluation directory
|
||||
├── images # directory of images
|
||||
└── labels # directory of labels
|
||||
```
|
||||
|
||||
After configuring the JSON file for segmentation or enhancement,
|
||||
training can be initiated by running the following command line,
|
||||
similar to classification and reading-order model training:
|
||||
|
||||
```sh
|
||||
eynollah-training train with config_classification.json
|
||||
```
|
||||
|
||||
#### Binarization
|
||||
|
||||
### Ground truth format
|
||||
|
||||
Lables for each pixel are identified by a number. So if you have a
|
||||
binary case, ``n_classes`` should be set to ``2`` and labels should
|
||||
be ``0`` and ``1`` for each class and pixel.
|
||||
|
||||
In the case of multiclass, just set ``n_classes`` to the number of classes
|
||||
you have and the try to produce the labels by pixels set from ``0 , 1 ,2 .., n_classes-1``.
|
||||
The labels format should be png.
|
||||
Our lables are 3 channel png images but only information of first channel is used.
|
||||
If you have an image label with height and width of 10, for a binary case the first channel should look like this:
|
||||
|
||||
Label: [ [1, 0, 0, 1, 1, 0, 0, 1, 0, 0],
|
||||
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
|
||||
...,
|
||||
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
|
||||
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0] ]
|
||||
|
||||
This means that you have an image by `10*10*3` and `pixel[0,0]` belongs
|
||||
to class `1` and `pixel[0,1]` belongs to class `0`.
|
||||
|
||||
A small sample of training data for binarization experiment can be found here, [Training data sample](https://qurator-data.de/~vahid.rezanezhad/binarization_training_data_sample/), which contains images and lables folders.
|
||||
|
||||
|
||||
An example config json file for binarization can be like this:
|
||||
|
||||
```yaml
|
||||
{
|
||||
"backbone_type" : "transformer",
|
||||
"task": "binarization",
|
||||
"n_classes" : 2,
|
||||
"n_epochs" : 4,
|
||||
"input_height" : 224,
|
||||
"input_width" : 672,
|
||||
"weight_decay" : 1e-6,
|
||||
"n_batch" : 1,
|
||||
"learning_rate": 1e-4,
|
||||
"patches" : true,
|
||||
"pretraining" : true,
|
||||
"augmentation" : true,
|
||||
"flip_aug" : false,
|
||||
"blur_aug" : false,
|
||||
"scaling" : true,
|
||||
"degrading": false,
|
||||
"brightening": false,
|
||||
"binarization" : false,
|
||||
"scaling_bluring" : false,
|
||||
"scaling_binarization" : false,
|
||||
"scaling_flip" : false,
|
||||
"rotation": false,
|
||||
"rotation_not_90": false,
|
||||
"transformer_num_patches_xy": [7, 7],
|
||||
"transformer_patchsize_x": 3,
|
||||
"transformer_patchsize_y": 1,
|
||||
"transformer_projection_dim": 192,
|
||||
"transformer_mlp_head_units": [128, 64],
|
||||
"transformer_layers": 8,
|
||||
"transformer_num_heads": 4,
|
||||
"transformer_cnn_first": true,
|
||||
"blur_k" : ["blur","guass","median"],
|
||||
"scales" : [0.6, 0.7, 0.8, 0.9, 1.1, 1.2, 1.4],
|
||||
"brightness" : [1.3, 1.5, 1.7, 2],
|
||||
"degrade_scales" : [0.2, 0.4],
|
||||
"flip_index" : [0, 1, -1],
|
||||
"thetha" : [10, -10],
|
||||
"continue_training": false,
|
||||
"index_start" : 0,
|
||||
"dir_of_start_model" : " ",
|
||||
"weighted_loss": false,
|
||||
"is_loss_soft_dice": false,
|
||||
"data_is_provided": false,
|
||||
"dir_train": "./train",
|
||||
"dir_eval": "./eval",
|
||||
"dir_output": "./output"
|
||||
}
|
||||
```
|
||||
|
||||
#### Textline
|
||||
|
||||
```yaml
|
||||
{
|
||||
"backbone_type" : "nontransformer",
|
||||
"task": "segmentation",
|
||||
"n_classes" : 2,
|
||||
"n_epochs" : 4,
|
||||
"input_height" : 448,
|
||||
"input_width" : 224,
|
||||
"weight_decay" : 1e-6,
|
||||
"n_batch" : 1,
|
||||
"learning_rate": 1e-4,
|
||||
"patches" : true,
|
||||
"pretraining" : true,
|
||||
"augmentation" : true,
|
||||
"flip_aug" : false,
|
||||
"blur_aug" : false,
|
||||
"scaling" : true,
|
||||
"degrading": false,
|
||||
"brightening": false,
|
||||
"binarization" : false,
|
||||
"scaling_bluring" : false,
|
||||
"scaling_binarization" : false,
|
||||
"scaling_flip" : false,
|
||||
"rotation": false,
|
||||
"rotation_not_90": false,
|
||||
"blur_k" : ["blur","guass","median"],
|
||||
"scales" : [0.6, 0.7, 0.8, 0.9, 1.1, 1.2, 1.4],
|
||||
"brightness" : [1.3, 1.5, 1.7, 2],
|
||||
"degrade_scales" : [0.2, 0.4],
|
||||
"flip_index" : [0, 1, -1],
|
||||
"thetha" : [10, -10],
|
||||
"continue_training": false,
|
||||
"index_start" : 0,
|
||||
"dir_of_start_model" : " ",
|
||||
"weighted_loss": false,
|
||||
"is_loss_soft_dice": false,
|
||||
"data_is_provided": false,
|
||||
"dir_train": "./train",
|
||||
"dir_eval": "./eval",
|
||||
"dir_output": "./output"
|
||||
}
|
||||
```
|
||||
|
||||
#### Enhancement
|
||||
|
||||
```yaml
|
||||
{
|
||||
"backbone_type" : "nontransformer",
|
||||
"task": "enhancement",
|
||||
"n_classes" : 3,
|
||||
"n_epochs" : 4,
|
||||
"input_height" : 448,
|
||||
"input_width" : 224,
|
||||
"weight_decay" : 1e-6,
|
||||
"n_batch" : 4,
|
||||
"learning_rate": 1e-4,
|
||||
"patches" : true,
|
||||
"pretraining" : true,
|
||||
"augmentation" : true,
|
||||
"flip_aug" : false,
|
||||
"blur_aug" : false,
|
||||
"scaling" : true,
|
||||
"degrading": false,
|
||||
"brightening": false,
|
||||
"binarization" : false,
|
||||
"scaling_bluring" : false,
|
||||
"scaling_binarization" : false,
|
||||
"scaling_flip" : false,
|
||||
"rotation": false,
|
||||
"rotation_not_90": false,
|
||||
"blur_k" : ["blur","guass","median"],
|
||||
"scales" : [0.6, 0.7, 0.8, 0.9, 1.1, 1.2, 1.4],
|
||||
"brightness" : [1.3, 1.5, 1.7, 2],
|
||||
"degrade_scales" : [0.2, 0.4],
|
||||
"flip_index" : [0, 1, -1],
|
||||
"thetha" : [10, -10],
|
||||
"continue_training": false,
|
||||
"index_start" : 0,
|
||||
"dir_of_start_model" : " ",
|
||||
"weighted_loss": false,
|
||||
"is_loss_soft_dice": false,
|
||||
"data_is_provided": false,
|
||||
"dir_train": "./train",
|
||||
"dir_eval": "./eval",
|
||||
"dir_output": "./output"
|
||||
}
|
||||
```
|
||||
|
||||
It's important to mention that the value of n_classes for enhancement should be 3, as the model's output is a 3-channel
|
||||
image.
|
||||
|
||||
#### Page extraction
|
||||
|
||||
```yaml
|
||||
{
|
||||
"backbone_type" : "nontransformer",
|
||||
"task": "segmentation",
|
||||
"n_classes" : 2,
|
||||
"n_epochs" : 4,
|
||||
"input_height" : 448,
|
||||
"input_width" : 224,
|
||||
"weight_decay" : 1e-6,
|
||||
"n_batch" : 1,
|
||||
"learning_rate": 1e-4,
|
||||
"patches" : false,
|
||||
"pretraining" : true,
|
||||
"augmentation" : false,
|
||||
"flip_aug" : false,
|
||||
"blur_aug" : false,
|
||||
"scaling" : true,
|
||||
"degrading": false,
|
||||
"brightening": false,
|
||||
"binarization" : false,
|
||||
"scaling_bluring" : false,
|
||||
"scaling_binarization" : false,
|
||||
"scaling_flip" : false,
|
||||
"rotation": false,
|
||||
"rotation_not_90": false,
|
||||
"blur_k" : ["blur","guass","median"],
|
||||
"scales" : [0.6, 0.7, 0.8, 0.9, 1.1, 1.2, 1.4],
|
||||
"brightness" : [1.3, 1.5, 1.7, 2],
|
||||
"degrade_scales" : [0.2, 0.4],
|
||||
"flip_index" : [0, 1, -1],
|
||||
"thetha" : [10, -10],
|
||||
"continue_training": false,
|
||||
"index_start" : 0,
|
||||
"dir_of_start_model" : " ",
|
||||
"weighted_loss": false,
|
||||
"is_loss_soft_dice": false,
|
||||
"data_is_provided": false,
|
||||
"dir_train": "./train",
|
||||
"dir_eval": "./eval",
|
||||
"dir_output": "./output"
|
||||
}
|
||||
```
|
||||
|
||||
For page segmentation (or print space or border segmentation), the model needs to view the input image in its
|
||||
entirety,hence the patches parameter should be set to false.
|
||||
|
||||
#### layout segmentation
|
||||
|
||||
An example config json file for layout segmentation with 5 classes (including background) can be like this:
|
||||
|
||||
```yaml
|
||||
{
|
||||
"backbone_type" : "transformer",
|
||||
"task": "segmentation",
|
||||
"n_classes" : 5,
|
||||
"n_epochs" : 4,
|
||||
"input_height" : 448,
|
||||
"input_width" : 224,
|
||||
"weight_decay" : 1e-6,
|
||||
"n_batch" : 1,
|
||||
"learning_rate": 1e-4,
|
||||
"patches" : true,
|
||||
"pretraining" : true,
|
||||
"augmentation" : true,
|
||||
"flip_aug" : false,
|
||||
"blur_aug" : false,
|
||||
"scaling" : true,
|
||||
"degrading": false,
|
||||
"brightening": false,
|
||||
"binarization" : false,
|
||||
"scaling_bluring" : false,
|
||||
"scaling_binarization" : false,
|
||||
"scaling_flip" : false,
|
||||
"rotation": false,
|
||||
"rotation_not_90": false,
|
||||
"transformer_num_patches_xy": [7, 14],
|
||||
"transformer_patchsize_x": 1,
|
||||
"transformer_patchsize_y": 1,
|
||||
"transformer_projection_dim": 64,
|
||||
"transformer_mlp_head_units": [128, 64],
|
||||
"transformer_layers": 8,
|
||||
"transformer_num_heads": 4,
|
||||
"transformer_cnn_first": true,
|
||||
"blur_k" : ["blur","guass","median"],
|
||||
"scales" : [0.6, 0.7, 0.8, 0.9, 1.1, 1.2, 1.4],
|
||||
"brightness" : [1.3, 1.5, 1.7, 2],
|
||||
"degrade_scales" : [0.2, 0.4],
|
||||
"flip_index" : [0, 1, -1],
|
||||
"thetha" : [10, -10],
|
||||
"continue_training": false,
|
||||
"index_start" : 0,
|
||||
"dir_of_start_model" : " ",
|
||||
"weighted_loss": false,
|
||||
"is_loss_soft_dice": false,
|
||||
"data_is_provided": false,
|
||||
"dir_train": "./train",
|
||||
"dir_eval": "./eval",
|
||||
"dir_output": "./output"
|
||||
}
|
||||
```
|
||||
## Inference with the trained model
|
||||
|
||||
### classification
|
||||
|
||||
For conducting inference with a trained model, you simply need to execute the following command line, specifying the
|
||||
directory of the model and the image on which to perform inference:
|
||||
|
||||
```sh
|
||||
eynollah-training inference -m "model dir" -i "image"
|
||||
```
|
||||
|
||||
This will straightforwardly return the class of the image.
|
||||
|
||||
### machine based reading order
|
||||
|
||||
To infer the reading order using a reading order model, we need a PAGE XML file containing layout information but
|
||||
without the reading order. We simply need to provide the model directory, the XML file, and the output directory. The
|
||||
new XML file with the added reading order will be written to the output directory with the same name. We need to run:
|
||||
|
||||
```sh
|
||||
eynollah-training inference \
|
||||
-m "model dir" \
|
||||
-xml "page xml file" \
|
||||
-o "output dir to write new xml with reading order"
|
||||
```
|
||||
|
||||
### Segmentation (Textline, Binarization, Page extraction and layout) and enhancement
|
||||
|
||||
For conducting inference with a trained model for segmentation and enhancement you need to run the following command line:
|
||||
|
||||
```sh
|
||||
eynollah-training inference \
|
||||
-m "model dir" \
|
||||
-i "image" \
|
||||
-p \
|
||||
-s "output image"
|
||||
```
|
||||
|
||||
Note that in the case of page extraction the -p flag is not needed.
|
||||
|
||||
For segmentation or binarization tasks, if a ground truth (GT) label is available, the IoU evaluation metric can be
|
||||
calculated for the output. To do this, you need to provide the GT label using the argument -gt.
|
||||
|
|
@ -1,92 +0,0 @@
|
|||
# Usage documentation
|
||||
The command-line interface can be called like this:
|
||||
|
||||
```sh
|
||||
eynollah \
|
||||
-i <single image file> | -di <directory containing image files> \
|
||||
-o <output directory> \
|
||||
-m <directory containing model files> \
|
||||
[OPTIONS]
|
||||
```
|
||||
|
||||
## Processing options
|
||||
The following options can be used to further configure the processing:
|
||||
|
||||
| option | description |
|
||||
|-------------------|:-------------------------------------------------------------------------------|
|
||||
| `-fl` | full layout analysis including all steps and segmentation classes |
|
||||
| `-light` | lighter and faster but simpler method for main region detection and deskewing |
|
||||
| `-tab` | apply table detection |
|
||||
| `-ae` | apply enhancement (the resulting image is saved to the output directory) |
|
||||
| `-as` | apply scaling |
|
||||
| `-cl` | apply contour detection for curved text lines instead of bounding boxes |
|
||||
| `-ib` | apply binarization (the resulting image is saved to the output directory) |
|
||||
| `-ep` | enable plotting (MUST always be used with `-sl`, `-sd`, `-sa`, `-si` or `-ae`) |
|
||||
| `-eoi` | extract only images to output directory (other processing will not be done) |
|
||||
| `-ho` | ignore headers for reading order dectection |
|
||||
| `-si <directory>` | save image regions detected to this directory |
|
||||
| `-sd <directory>` | save deskewed image to this directory |
|
||||
| `-sl <directory>` | save layout prediction as plot to this directory |
|
||||
| `-sp <directory>` | save cropped page image to this directory |
|
||||
| `-sa <directory>` | save all (plot, enhanced/binary image, layout) to this directory |
|
||||
|
||||
If no option is set, the tool performs detection of main regions (background, text, images, separators and marginals).
|
||||
|
||||
### `--full-layout` vs `--no-full-layout`
|
||||
|
||||
Here are the difference in elements detected depending on the `--full-layout`/`--no-full-layout` command line flags:
|
||||
|
||||
| | `--full-layout` | `--no-full-layout` |
|
||||
|--------------------------|-----------------|--------------------|
|
||||
| reading order | x | x |
|
||||
| header regions | x | - |
|
||||
| text regions | x | x |
|
||||
| text regions / text line | x | x |
|
||||
| drop-capitals | x | - |
|
||||
| marginals | x | x |
|
||||
| marginals / text line | x | x |
|
||||
| image region | x | x |
|
||||
|
||||
## Use as OCR-D processor
|
||||
Eynollah ships with a CLI interface to be used as [OCR-D](https://ocr-d.de) processor that is described in
|
||||
[`ocrd-tool.json`](https://github.com/qurator-spk/eynollah/tree/main/src/eynollah/ocrd-tool.json).
|
||||
|
||||
The source image file group with (preferably) RGB images should be used as input for Eynollah like this:
|
||||
|
||||
```
|
||||
ocrd-eynollah-segment -I OCR-D-IMG -O SEG-LINE -P models
|
||||
```
|
||||
|
||||
Any image referenced by `@imageFilename` in PAGE-XML is passed on directly to Eynollah as a processor, so that e.g.
|
||||
|
||||
```
|
||||
ocrd-eynollah-segment -I OCR-D-IMG-BIN -O SEG-LINE -P models
|
||||
```
|
||||
|
||||
uses the original (RGB) image despite any binarization that may have occured in previous OCR-D processing steps.
|
||||
|
||||
## Use with Docker
|
||||
TODO
|
||||
|
||||
## Hints
|
||||
* The best output quality is produced when RGB images are used as input rather than greyscale or binarized images.
|
||||
* If none of the parameters is set to `true`, the tool will perform a layout detection of main regions (background,
|
||||
text, images, separators and marginals). An advantage of this tool is that it tries to extract main text regions
|
||||
separately as much as possible.
|
||||
* If you set `-ae` (**a**llow image **e**nhancement) parameter to `true`, the tool will first check the ppi
|
||||
(pixel-per-inch) of the image and when it is less than 300, the tool will resize it and only then image enhancement will
|
||||
occur. Image enhancement can also take place without this option, but by setting this option to `true`, the layout xml
|
||||
data (e.g. coordinates) will be based on the resized and enhanced image instead of the original image.
|
||||
* For some documents, while the quality is good, their scale is very large, and the performance of tool decreases. In
|
||||
such cases you can set `-as` (**a**llow **s**caling) to `true`. With this option enabled, the tool will try to rescale
|
||||
the image and only then the layout detection process will begin.
|
||||
* If you care about drop capitals (initials) and headings, you can set `-fl` (**f**ull **l**ayout) to `true`. With this
|
||||
setting, the tool can currently distinguish 7 document layout classes/elements.
|
||||
* In cases where the document includes curved headers or curved lines, rectangular bounding boxes for textlines will not
|
||||
be a great option. In such cases it is strongly recommended setting the flag `-cl` (**c**urved **l**ines) to `true` to
|
||||
find contours of curved lines instead of rectangular bounding boxes. Be advised that enabling this option increases the
|
||||
processing time of the tool.
|
||||
* To crop and save image regions inside the document, set the parameter `-si` (**s**ave **i**mages) to true and provide
|
||||
a directory path to store the extracted images.
|
||||
* To extract only images from a document, set the parameter `-eoi` (**e**xtract **o**nly **i**mages). Choosing this
|
||||
option disables any other processing. To save the cropped images add `-ep` and `-si`.
|
||||
|
|
@ -1 +1 @@
|
|||
src/eynollah/ocrd-tool.json
|
||||
qurator/eynollah/ocrd-tool.json
|
||||
|
|
@ -1,85 +0,0 @@
|
|||
[build-system]
|
||||
requires = ["setuptools>=61.0", "wheel", "setuptools-ocrd"]
|
||||
|
||||
[project]
|
||||
name = "eynollah"
|
||||
authors = [
|
||||
{name = "Vahid Rezanezhad"},
|
||||
{name = "Staatsbibliothek zu Berlin - Preußischer Kulturbesitz"},
|
||||
]
|
||||
description = "Document Layout Analysis"
|
||||
readme = "README.md"
|
||||
license.file = "LICENSE"
|
||||
requires-python = ">=3.8"
|
||||
keywords = [
|
||||
"document layout analysis",
|
||||
"image segmentation",
|
||||
"binarization",
|
||||
"optical character recognition"
|
||||
]
|
||||
|
||||
dynamic = [
|
||||
"dependencies",
|
||||
"optional-dependencies",
|
||||
"version"
|
||||
]
|
||||
|
||||
classifiers = [
|
||||
"Development Status :: 4 - Beta",
|
||||
"Environment :: Console",
|
||||
"Intended Audience :: Science/Research",
|
||||
"License :: OSI Approved :: Apache Software License",
|
||||
"Programming Language :: Python :: 3",
|
||||
"Programming Language :: Python :: 3.8",
|
||||
"Programming Language :: Python :: 3.9",
|
||||
"Programming Language :: Python :: 3.10",
|
||||
"Programming Language :: Python :: 3.11",
|
||||
"Programming Language :: Python :: 3 :: Only",
|
||||
"Topic :: Scientific/Engineering :: Image Processing",
|
||||
]
|
||||
|
||||
[project.scripts]
|
||||
eynollah = "eynollah.cli:main"
|
||||
eynollah-training = "eynollah.training.cli:main"
|
||||
ocrd-eynollah-segment = "eynollah.ocrd_cli:main"
|
||||
ocrd-sbb-binarize = "eynollah.ocrd_cli_binarization:main"
|
||||
|
||||
[project.urls]
|
||||
Homepage = "https://github.com/qurator-spk/eynollah"
|
||||
Repository = "https://github.com/qurator-spk/eynollah.git"
|
||||
|
||||
[tool.setuptools.dynamic]
|
||||
dependencies = {file = ["requirements.txt"]}
|
||||
optional-dependencies.test = {file = ["requirements-test.txt"]}
|
||||
optional-dependencies.OCR = {file = ["requirements-ocr.txt"]}
|
||||
optional-dependencies.plotting = {file = ["requirements-plotting.txt"]}
|
||||
optional-dependencies.training = {file = ["requirements-training.txt"]}
|
||||
|
||||
[tool.setuptools.packages.find]
|
||||
where = ["src"]
|
||||
|
||||
[tool.setuptools.package-data]
|
||||
"*" = ["*.json", '*.yml', '*.xml', '*.xsd', '*.ttf']
|
||||
|
||||
[tool.coverage.run]
|
||||
branch = true
|
||||
source = ["eynollah"]
|
||||
|
||||
[tool.ruff]
|
||||
line-length = 120
|
||||
|
||||
[tool.ruff.lint]
|
||||
ignore = [
|
||||
# disable unused imports
|
||||
"F401",
|
||||
# disable import order
|
||||
"E402",
|
||||
# disable unused variables
|
||||
"F841",
|
||||
# disable bare except
|
||||
"E722",
|
||||
]
|
||||
|
||||
[tool.ruff.format]
|
||||
quote-style = "preserve"
|
||||
|
||||
1
qurator/__init__.py
Normal file
1
qurator/__init__.py
Normal file
|
|
@ -0,0 +1 @@
|
|||
__import__("pkg_resources").declare_namespace(__name__)
|
||||
1
qurator/eynollah/__init__.py
Normal file
1
qurator/eynollah/__init__.py
Normal file
|
|
@ -0,0 +1 @@
|
|||
|
||||
146
qurator/eynollah/cli.py
Normal file
146
qurator/eynollah/cli.py
Normal file
|
|
@ -0,0 +1,146 @@
|
|||
import sys
|
||||
import click
|
||||
from ocrd_utils import initLogging, setOverrideLogLevel
|
||||
from qurator.eynollah.eynollah import Eynollah
|
||||
|
||||
|
||||
@click.command()
|
||||
@click.option(
|
||||
"--image",
|
||||
"-i",
|
||||
help="image filename",
|
||||
type=click.Path(exists=True, dir_okay=False),
|
||||
required=True,
|
||||
)
|
||||
@click.option(
|
||||
"--out",
|
||||
"-o",
|
||||
help="directory to write output xml data",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
required=True,
|
||||
)
|
||||
@click.option(
|
||||
"--model",
|
||||
"-m",
|
||||
help="directory of models",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
)
|
||||
@click.option(
|
||||
"--save_images",
|
||||
"-si",
|
||||
help="if a directory is given, images in documents will be cropped and saved there",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
)
|
||||
@click.option(
|
||||
"--save_layout",
|
||||
"-sl",
|
||||
help="if a directory is given, plot of layout will be saved there",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
)
|
||||
@click.option(
|
||||
"--save_deskewed",
|
||||
"-sd",
|
||||
help="if a directory is given, deskewed image will be saved there",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
)
|
||||
@click.option(
|
||||
"--save_all",
|
||||
"-sa",
|
||||
help="if a directory is given, all plots needed for documentation will be saved there",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
)
|
||||
@click.option(
|
||||
"--enable-plotting/--disable-plotting",
|
||||
"-ep/-noep",
|
||||
is_flag=True,
|
||||
help="If set, will plot intermediary files and images",
|
||||
)
|
||||
@click.option(
|
||||
"--allow-enhancement/--no-allow-enhancement",
|
||||
"-ae/-noae",
|
||||
is_flag=True,
|
||||
help="if this parameter set to true, this tool would check that input image need resizing and enhancement or not. If so output of resized and enhanced image and corresponding layout data will be written in out directory",
|
||||
)
|
||||
@click.option(
|
||||
"--curved-line/--no-curvedline",
|
||||
"-cl/-nocl",
|
||||
is_flag=True,
|
||||
help="if this parameter set to true, this tool will try to return contoure of textlines instead of rectabgle bounding box of textline. This should be taken into account that with this option the tool need more time to do process.",
|
||||
)
|
||||
@click.option(
|
||||
"--full-layout/--no-full-layout",
|
||||
"-fl/-nofl",
|
||||
is_flag=True,
|
||||
help="if this parameter set to true, this tool will try to return all elements of layout.",
|
||||
)
|
||||
@click.option(
|
||||
"--input_binary/--input-RGB",
|
||||
"-ib/-irgb",
|
||||
is_flag=True,
|
||||
help="in general, eynollah uses RGB as input but if the input document is strongly dark, bright or for any other reason you can turn binarized input on. This option does not mean that you have to provide a binary image, otherwise this means that the tool itself will binarized the RGB input document.",
|
||||
)
|
||||
@click.option(
|
||||
"--allow_scaling/--no-allow-scaling",
|
||||
"-as/-noas",
|
||||
is_flag=True,
|
||||
help="if this parameter set to true, this tool would check the scale and if needed it will scale it to perform better layout detection",
|
||||
)
|
||||
@click.option(
|
||||
"--headers-off/--headers-on",
|
||||
"-ho/-noho",
|
||||
is_flag=True,
|
||||
help="if this parameter set to true, this tool would ignore headers role in reading order",
|
||||
)
|
||||
@click.option(
|
||||
"--log-level",
|
||||
"-l",
|
||||
type=click.Choice(['OFF', 'DEBUG', 'INFO', 'WARN', 'ERROR']),
|
||||
help="Override log level globally to this",
|
||||
)
|
||||
def main(
|
||||
image,
|
||||
out,
|
||||
model,
|
||||
save_images,
|
||||
save_layout,
|
||||
save_deskewed,
|
||||
save_all,
|
||||
enable_plotting,
|
||||
allow_enhancement,
|
||||
curved_line,
|
||||
full_layout,
|
||||
input_binary,
|
||||
allow_scaling,
|
||||
headers_off,
|
||||
log_level
|
||||
):
|
||||
if log_level:
|
||||
setOverrideLogLevel(log_level)
|
||||
initLogging()
|
||||
if not enable_plotting and (save_layout or save_deskewed or save_all or save_images):
|
||||
print("Error: You used one of -sl, -sd, -sa or -si but did not enable plotting with -ep")
|
||||
sys.exit(1)
|
||||
elif enable_plotting and not (save_layout or save_deskewed or save_all or save_images):
|
||||
print("Error: You used -ep to enable plotting but set none of -sl, -sd, -sa or -si")
|
||||
sys.exit(1)
|
||||
eynollah = Eynollah(
|
||||
image_filename=image,
|
||||
dir_out=out,
|
||||
dir_models=model,
|
||||
dir_of_cropped_images=save_images,
|
||||
dir_of_layout=save_layout,
|
||||
dir_of_deskewed=save_deskewed,
|
||||
dir_of_all=save_all,
|
||||
enable_plotting=enable_plotting,
|
||||
allow_enhancement=allow_enhancement,
|
||||
curved_line=curved_line,
|
||||
full_layout=full_layout,
|
||||
input_binary=input_binary,
|
||||
allow_scaling=allow_scaling,
|
||||
headers_off=headers_off,
|
||||
)
|
||||
pcgts = eynollah.run()
|
||||
eynollah.writer.write_pagexml(pcgts)
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
2086
qurator/eynollah/eynollah.py
Normal file
2086
qurator/eynollah/eynollah.py
Normal file
File diff suppressed because it is too large
Load diff
49
qurator/eynollah/ocrd-tool.json
Normal file
49
qurator/eynollah/ocrd-tool.json
Normal file
|
|
@ -0,0 +1,49 @@
|
|||
{
|
||||
"version": "0.0.8",
|
||||
"git_url": "https://github.com/qurator-spk/eynollah",
|
||||
"tools": {
|
||||
"ocrd-eynollah-segment": {
|
||||
"executable": "ocrd-eynollah-segment",
|
||||
"categories": ["Layout analysis"],
|
||||
"description": "Segment page into regions and lines and do reading order detection with eynollah",
|
||||
"input_file_grp": ["OCR-D-IMG", "OCR-D-SEG-PAGE", "OCR-D-GT-SEG-PAGE"],
|
||||
"output_file_grp": ["OCR-D-SEG-LINE"],
|
||||
"steps": ["layout/segmentation/region", "layout/segmentation/line"],
|
||||
"parameters": {
|
||||
"models": {
|
||||
"type": "string",
|
||||
"format": "file",
|
||||
"cacheable": true,
|
||||
"description": "Path to directory containing models to be used (See https://qurator-data.de/eynollah)",
|
||||
"required": true
|
||||
},
|
||||
"dpi": {
|
||||
"type": "number",
|
||||
"format": "float",
|
||||
"description": "pixel density in dots per inch (overrides any meta-data in the images); ignored if <= 0 (with fall-back 230)",
|
||||
"default": 0
|
||||
},
|
||||
"full_layout": {
|
||||
"type": "boolean",
|
||||
"default": true,
|
||||
"description": "Try to detect all element subtypes, including drop-caps and headings"
|
||||
},
|
||||
"curved_line": {
|
||||
"type": "boolean",
|
||||
"default": false,
|
||||
"description": "try to return contour of textlines instead of just rectangle bounding box. Needs more processing time"
|
||||
},
|
||||
"allow_scaling": {
|
||||
"type": "boolean",
|
||||
"default": false,
|
||||
"description": "check the resolution against the number of detected columns and if needed, scale the image up or down during layout detection (heuristic to improve quality and performance)"
|
||||
},
|
||||
"headers_off": {
|
||||
"type": "boolean",
|
||||
"default": false,
|
||||
"description": "ignore the special role of headings during reading order detection"
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
|
@ -1,6 +1,3 @@
|
|||
# NOTE: For predictable order of imports of torch/shapely/tensorflow
|
||||
# this must be the first import of the CLI!
|
||||
from .eynollah_imports import imported_libs
|
||||
from .processor import EynollahProcessor
|
||||
from click import command
|
||||
from ocrd.decorators import ocrd_cli_options, ocrd_cli_wrap_processor
|
||||
|
|
@ -1,8 +1,5 @@
|
|||
try:
|
||||
import matplotlib.pyplot as plt
|
||||
import matplotlib.patches as mpatches
|
||||
except ImportError:
|
||||
plt = mpatches = None
|
||||
import numpy as np
|
||||
import os.path
|
||||
import cv2
|
||||
|
|
@ -12,7 +9,7 @@ from .utils import crop_image_inside_box
|
|||
from .utils.rotate import rotate_image_different
|
||||
from .utils.resize import resize_image
|
||||
|
||||
class EynollahPlotter:
|
||||
class EynollahPlotter():
|
||||
"""
|
||||
Class collecting all the plotting and image writing methods
|
||||
"""
|
||||
|
|
@ -20,21 +17,26 @@ class EynollahPlotter:
|
|||
def __init__(
|
||||
self,
|
||||
*,
|
||||
dir_out,
|
||||
dir_of_all,
|
||||
dir_save_page,
|
||||
dir_of_deskewed,
|
||||
dir_of_layout,
|
||||
dir_of_cropped_images,
|
||||
image_filename_stem,
|
||||
image_org=None,
|
||||
scale_x=1,
|
||||
scale_y=1,
|
||||
):
|
||||
self.dir_out = dir_out
|
||||
self.dir_of_all = dir_of_all
|
||||
self.dir_save_page = dir_save_page
|
||||
self.dir_of_layout = dir_of_layout
|
||||
self.dir_of_cropped_images = dir_of_cropped_images
|
||||
self.dir_of_deskewed = dir_of_deskewed
|
||||
self.image_filename_stem = image_filename_stem
|
||||
# XXX TODO hacky these cannot be set at init time
|
||||
self.image_org = image_org
|
||||
self.scale_x = scale_x
|
||||
self.scale_y = scale_y
|
||||
|
||||
def save_plot_of_layout_main(self, text_regions_p, image_page, name=None):
|
||||
def save_plot_of_layout_main(self, text_regions_p, image_page):
|
||||
if self.dir_of_layout is not None:
|
||||
values = np.unique(text_regions_p[:, :])
|
||||
# pixels=['Background' , 'Main text' , 'Heading' , 'Marginalia' ,'Drop capitals' , 'Images' , 'Seperators' , 'Tables', 'Graphics']
|
||||
|
|
@ -46,10 +48,10 @@ class EynollahPlotter:
|
|||
colors = [im.cmap(im.norm(value)) for value in values]
|
||||
patches = [mpatches.Patch(color=colors[np.where(values == i)[0][0]], label="{l}".format(l=pixels[int(np.where(values_indexes == i)[0][0])])) for i in values]
|
||||
plt.legend(handles=patches, bbox_to_anchor=(1.05, 1), loc=2, borderaxespad=0.0, fontsize=40)
|
||||
plt.savefig(os.path.join(self.dir_of_layout,
|
||||
(name or "page") + "_layout_main.png"))
|
||||
plt.savefig(os.path.join(self.dir_of_layout, self.image_filename_stem + "_layout_main.png"))
|
||||
|
||||
def save_plot_of_layout_main_all(self, text_regions_p, image_page, name=None):
|
||||
|
||||
def save_plot_of_layout_main_all(self, text_regions_p, image_page):
|
||||
if self.dir_of_all is not None:
|
||||
values = np.unique(text_regions_p[:, :])
|
||||
# pixels=['Background' , 'Main text' , 'Heading' , 'Marginalia' ,'Drop capitals' , 'Images' , 'Seperators' , 'Tables', 'Graphics']
|
||||
|
|
@ -64,30 +66,28 @@ class EynollahPlotter:
|
|||
colors = [im.cmap(im.norm(value)) for value in values]
|
||||
patches = [mpatches.Patch(color=colors[np.where(values == i)[0][0]], label="{l}".format(l=pixels[int(np.where(values_indexes == i)[0][0])])) for i in values]
|
||||
plt.legend(handles=patches, bbox_to_anchor=(1.05, 1), loc=2, borderaxespad=0.0, fontsize=60)
|
||||
plt.savefig(os.path.join(self.dir_of_all,
|
||||
(name or "page") + "_layout_main_and_page.png"))
|
||||
plt.savefig(os.path.join(self.dir_of_all, self.image_filename_stem + "_layout_main_and_page.png"))
|
||||
|
||||
def save_plot_of_layout(self, text_regions_p, image_page, name=None):
|
||||
def save_plot_of_layout(self, text_regions_p, image_page):
|
||||
if self.dir_of_layout is not None:
|
||||
values = np.unique(text_regions_p[:, :])
|
||||
# pixels=['Background' , 'Main text' , 'Heading' , 'Marginalia' ,'Drop capitals' , 'Images' , 'Seperators' , 'Tables', 'Graphics']
|
||||
pixels = ["Background", "Main text", "Header", "Marginalia", "Drop capital", "Image", "Separator", "Tables"]
|
||||
values_indexes = [0, 1, 2, 8, 4, 5, 6, 10]
|
||||
pixels = ["Background", "Main text", "Header", "Marginalia", "Drop capital", "Image", "Separator"]
|
||||
values_indexes = [0, 1, 2, 8, 4, 5, 6]
|
||||
plt.figure(figsize=(40, 40))
|
||||
plt.rcParams["font.size"] = "40"
|
||||
im = plt.imshow(text_regions_p[:, :])
|
||||
colors = [im.cmap(im.norm(value)) for value in values]
|
||||
patches = [mpatches.Patch(color=colors[np.where(values == i)[0][0]], label="{l}".format(l=pixels[int(np.where(values_indexes == i)[0][0])])) for i in values]
|
||||
plt.legend(handles=patches, bbox_to_anchor=(1.05, 1), loc=2, borderaxespad=0.0, fontsize=40)
|
||||
plt.savefig(os.path.join(self.dir_of_layout,
|
||||
(name or "page") + "_layout.png"))
|
||||
plt.savefig(os.path.join(self.dir_of_layout, self.image_filename_stem + "_layout.png"))
|
||||
|
||||
def save_plot_of_layout_all(self, text_regions_p, image_page, name=None):
|
||||
def save_plot_of_layout_all(self, text_regions_p, image_page):
|
||||
if self.dir_of_all is not None:
|
||||
values = np.unique(text_regions_p[:, :])
|
||||
# pixels=['Background' , 'Main text' , 'Heading' , 'Marginalia' ,'Drop capitals' , 'Images' , 'Seperators' , 'Tables', 'Graphics']
|
||||
pixels = ["Background", "Main text", "Header", "Marginalia", "Drop capital", "Image", "Separator", "Tables"]
|
||||
values_indexes = [0, 1, 2, 8, 4, 5, 6, 10]
|
||||
pixels = ["Background", "Main text", "Header", "Marginalia", "Drop capital", "Image", "Separator"]
|
||||
values_indexes = [0, 1, 2, 8, 4, 5, 6]
|
||||
plt.figure(figsize=(80, 40))
|
||||
plt.rcParams["font.size"] = "40"
|
||||
plt.subplot(1, 2, 1)
|
||||
|
|
@ -97,10 +97,9 @@ class EynollahPlotter:
|
|||
colors = [im.cmap(im.norm(value)) for value in values]
|
||||
patches = [mpatches.Patch(color=colors[np.where(values == i)[0][0]], label="{l}".format(l=pixels[int(np.where(values_indexes == i)[0][0])])) for i in values]
|
||||
plt.legend(handles=patches, bbox_to_anchor=(1.05, 1), loc=2, borderaxespad=0.0, fontsize=60)
|
||||
plt.savefig(os.path.join(self.dir_of_all,
|
||||
(name or "page") + "_layout_and_page.png"))
|
||||
plt.savefig(os.path.join(self.dir_of_all, self.image_filename_stem + "_layout_and_page.png"))
|
||||
|
||||
def save_plot_of_textlines(self, textline_mask_tot_ea, image_page, name=None):
|
||||
def save_plot_of_textlines(self, textline_mask_tot_ea, image_page):
|
||||
if self.dir_of_all is not None:
|
||||
values = np.unique(textline_mask_tot_ea[:, :])
|
||||
pixels = ["Background", "Textlines"]
|
||||
|
|
@ -114,31 +113,20 @@ class EynollahPlotter:
|
|||
colors = [im.cmap(im.norm(value)) for value in values]
|
||||
patches = [mpatches.Patch(color=colors[np.where(values == i)[0][0]], label="{l}".format(l=pixels[int(np.where(values_indexes == i)[0][0])])) for i in values]
|
||||
plt.legend(handles=patches, bbox_to_anchor=(1.05, 1), loc=2, borderaxespad=0.0, fontsize=60)
|
||||
plt.savefig(os.path.join(self.dir_of_all,
|
||||
(name or "page") + "_textline_and_page.png"))
|
||||
plt.savefig(os.path.join(self.dir_of_all, self.image_filename_stem + "_textline_and_page.png"))
|
||||
|
||||
def save_deskewed_image(self, slope_deskew, image_org, name=None):
|
||||
def save_deskewed_image(self, slope_deskew):
|
||||
if self.dir_of_all is not None:
|
||||
cv2.imwrite(os.path.join(self.dir_of_all,
|
||||
(name or "page") + "_org.png"), image_org)
|
||||
cv2.imwrite(os.path.join(self.dir_of_all, self.image_filename_stem + "_org.png"), self.image_org)
|
||||
if self.dir_of_deskewed is not None:
|
||||
img_rotated = rotate_image_different(image_org, slope_deskew)
|
||||
cv2.imwrite(os.path.join(self.dir_of_deskewed,
|
||||
(name or "page") + "_deskewed.png"), img_rotated)
|
||||
img_rotated = rotate_image_different(self.image_org, slope_deskew)
|
||||
cv2.imwrite(os.path.join(self.dir_of_deskewed, self.image_filename_stem + "_deskewed.png"), img_rotated)
|
||||
|
||||
def save_page_image(self, image_page, name=None):
|
||||
def save_page_image(self, image_page):
|
||||
if self.dir_of_all is not None:
|
||||
cv2.imwrite(os.path.join(self.dir_of_all,
|
||||
(name or "page") + "_page.png"), image_page)
|
||||
if self.dir_save_page is not None:
|
||||
cv2.imwrite(os.path.join(self.dir_save_page,
|
||||
(name or "page") + "_page.png"), image_page)
|
||||
cv2.imwrite(os.path.join(self.dir_of_all, self.image_filename_stem + "_page.png"), image_page)
|
||||
|
||||
def save_enhanced_image(self, img_res, name=None):
|
||||
cv2.imwrite(os.path.join(self.dir_out,
|
||||
(name or "page") + "_enhanced.png"), img_res)
|
||||
|
||||
def save_plot_of_textline_density(self, img_patch_org, name=None):
|
||||
def save_plot_of_textline_density(self, img_patch_org):
|
||||
if self.dir_of_all is not None:
|
||||
plt.figure(figsize=(80,40))
|
||||
plt.rcParams['font.size']='50'
|
||||
|
|
@ -150,10 +138,9 @@ class EynollahPlotter:
|
|||
plt.ylabel('Height',fontsize=60)
|
||||
plt.yticks([0,len(gaussian_filter1d(img_patch_org.sum(axis=1), 3))])
|
||||
plt.gca().invert_yaxis()
|
||||
plt.savefig(os.path.join(self.dir_of_all,
|
||||
(name or "page") + '_density_of_textline.png'))
|
||||
plt.savefig(os.path.join(self.dir_of_all, self.image_filename_stem+'_density_of_textline.png'))
|
||||
|
||||
def save_plot_of_rotation_angle(self, angels, var_res, name=None):
|
||||
def save_plot_of_rotation_angle(self, angels, var_res):
|
||||
if self.dir_of_all is not None:
|
||||
plt.figure(figsize=(60,30))
|
||||
plt.rcParams['font.size']='50'
|
||||
|
|
@ -162,20 +149,19 @@ class EynollahPlotter:
|
|||
plt.ylabel('variance of sum of rotated textline in direction of x axis',fontsize=50)
|
||||
plt.plot(angels[np.argmax(var_res)],var_res[np.argmax(np.array(var_res))] ,'*',markersize=50,label='Angle of deskewing=' +str("{:.2f}".format(angels[np.argmax(var_res)]))+r'$\degree$')
|
||||
plt.legend(loc='best')
|
||||
plt.savefig(os.path.join(self.dir_of_all,
|
||||
(name or "page") + '_rotation_angle.png'))
|
||||
plt.savefig(os.path.join(self.dir_of_all, self.image_filename_stem+'_rotation_angle.png'))
|
||||
|
||||
def write_images_into_directory(self, img_contours, image_page, scale_x=1.0, scale_y=1.0, name=None):
|
||||
def write_images_into_directory(self, img_contours, image_page):
|
||||
if self.dir_of_cropped_images is not None:
|
||||
index = 0
|
||||
for cont_ind in img_contours:
|
||||
x, y, w, h = cv2.boundingRect(cont_ind)
|
||||
box = [x, y, w, h]
|
||||
image, _ = crop_image_inside_box(box, image_page)
|
||||
image = resize_image(image,
|
||||
int(image.shape[0] / scale_y),
|
||||
int(image.shape[1] / scale_x))
|
||||
cv2.imwrite(os.path.join(self.dir_of_cropped_images,
|
||||
(name or "page") + f"_{index:03d}.jpg"), image)
|
||||
croped_page, page_coord = crop_image_inside_box(box, image_page)
|
||||
|
||||
croped_page = resize_image(croped_page, int(croped_page.shape[0] / self.scale_y), int(croped_page.shape[1] / self.scale_x))
|
||||
|
||||
path = os.path.join(self.dir_of_cropped_images, self.image_filename_stem + "_" + str(index) + ".jpg")
|
||||
cv2.imwrite(path, croped_page)
|
||||
index += 1
|
||||
|
||||
67
qurator/eynollah/processor.py
Normal file
67
qurator/eynollah/processor.py
Normal file
|
|
@ -0,0 +1,67 @@
|
|||
from json import loads
|
||||
from pkg_resources import resource_string
|
||||
from tempfile import NamedTemporaryFile
|
||||
from pathlib import Path
|
||||
from os.path import join
|
||||
|
||||
from PIL import Image
|
||||
|
||||
from ocrd import Processor
|
||||
from ocrd_modelfactory import page_from_file, exif_from_filename
|
||||
from ocrd_models import OcrdFile, OcrdExif
|
||||
from ocrd_models.ocrd_page import to_xml
|
||||
from ocrd_utils import (
|
||||
getLogger,
|
||||
MIMETYPE_PAGE,
|
||||
assert_file_grp_cardinality,
|
||||
make_file_id
|
||||
)
|
||||
|
||||
from .eynollah import Eynollah
|
||||
from .utils.pil_cv2 import pil2cv
|
||||
|
||||
OCRD_TOOL = loads(resource_string(__name__, 'ocrd-tool.json').decode('utf8'))
|
||||
|
||||
class EynollahProcessor(Processor):
|
||||
|
||||
def __init__(self, *args, **kwargs):
|
||||
kwargs['ocrd_tool'] = OCRD_TOOL['tools']['ocrd-eynollah-segment']
|
||||
kwargs['version'] = OCRD_TOOL['version']
|
||||
super().__init__(*args, **kwargs)
|
||||
|
||||
def process(self):
|
||||
LOG = getLogger('eynollah')
|
||||
assert_file_grp_cardinality(self.input_file_grp, 1)
|
||||
assert_file_grp_cardinality(self.output_file_grp, 1)
|
||||
for n, input_file in enumerate(self.input_files):
|
||||
page_id = input_file.pageId or input_file.ID
|
||||
LOG.info("INPUT FILE %s (%d/%d) ", page_id, n + 1, len(self.input_files))
|
||||
pcgts = page_from_file(self.workspace.download_file(input_file))
|
||||
LOG.debug('width %s height %s', pcgts.get_Page().imageWidth, pcgts.get_Page().imageHeight)
|
||||
self.add_metadata(pcgts)
|
||||
page = pcgts.get_Page()
|
||||
# XXX loses DPI information
|
||||
# page_image, _, _ = self.workspace.image_from_page(page, page_id, feature_filter='binarized')
|
||||
image_filename = self.workspace.download_file(next(self.workspace.mets.find_files(url=page.imageFilename))).local_filename
|
||||
eynollah_kwargs = {
|
||||
'dir_models': self.resolve_resource(self.parameter['models']),
|
||||
'allow_enhancement': False,
|
||||
'curved_line': self.parameter['curved_line'],
|
||||
'full_layout': self.parameter['full_layout'],
|
||||
'allow_scaling': self.parameter['allow_scaling'],
|
||||
'headers_off': self.parameter['headers_off'],
|
||||
'override_dpi': self.parameter['dpi'],
|
||||
'logger': LOG,
|
||||
'pcgts': pcgts,
|
||||
'image_filename': image_filename
|
||||
}
|
||||
Eynollah(**eynollah_kwargs).run()
|
||||
file_id = make_file_id(input_file, self.output_file_grp)
|
||||
pcgts.set_pcGtsId(file_id)
|
||||
self.workspace.add_file(
|
||||
ID=file_id,
|
||||
file_grp=self.output_file_grp,
|
||||
pageId=page_id,
|
||||
mimetype=MIMETYPE_PAGE,
|
||||
local_filename=join(self.output_file_grp, file_id) + '.xml',
|
||||
content=to_xml(pcgts))
|
||||
2240
qurator/eynollah/utils/__init__.py
Normal file
2240
qurator/eynollah/utils/__init__.py
Normal file
File diff suppressed because it is too large
Load diff
238
qurator/eynollah/utils/contour.py
Normal file
238
qurator/eynollah/utils/contour.py
Normal file
|
|
@ -0,0 +1,238 @@
|
|||
import cv2
|
||||
import numpy as np
|
||||
from shapely import geometry
|
||||
|
||||
from .rotate import rotate_image, rotation_image_new
|
||||
|
||||
def contours_in_same_horizon(cy_main_hor):
|
||||
X1 = np.zeros((len(cy_main_hor), len(cy_main_hor)))
|
||||
X2 = np.zeros((len(cy_main_hor), len(cy_main_hor)))
|
||||
|
||||
X1[0::1, :] = cy_main_hor[:]
|
||||
X2 = X1.T
|
||||
|
||||
X_dif = np.abs(X2 - X1)
|
||||
args_help = np.array(range(len(cy_main_hor)))
|
||||
all_args = []
|
||||
for i in range(len(cy_main_hor)):
|
||||
list_h = list(args_help[X_dif[i, :] <= 20])
|
||||
list_h.append(i)
|
||||
if len(list_h) > 1:
|
||||
all_args.append(list(set(list_h)))
|
||||
return np.unique(all_args)
|
||||
|
||||
def find_contours_mean_y_diff(contours_main):
|
||||
M_main = [cv2.moments(contours_main[j]) for j in range(len(contours_main))]
|
||||
cy_main = [(M_main[j]["m01"] / (M_main[j]["m00"] + 1e-32)) for j in range(len(M_main))]
|
||||
return np.mean(np.diff(np.sort(np.array(cy_main))))
|
||||
|
||||
|
||||
def get_text_region_boxes_by_given_contours(contours):
|
||||
|
||||
kernel = np.ones((5, 5), np.uint8)
|
||||
boxes = []
|
||||
contours_new = []
|
||||
for jj in range(len(contours)):
|
||||
x, y, w, h = cv2.boundingRect(contours[jj])
|
||||
|
||||
boxes.append([x, y, w, h])
|
||||
contours_new.append(contours[jj])
|
||||
|
||||
del contours
|
||||
return boxes, contours_new
|
||||
|
||||
def filter_contours_area_of_image(image, contours, hierarchy, max_area, min_area):
|
||||
found_polygons_early = list()
|
||||
jv = 0
|
||||
for c in contours:
|
||||
if len(c) < 3: # A polygon cannot have less than 3 points
|
||||
continue
|
||||
|
||||
polygon = geometry.Polygon([point[0] for point in c])
|
||||
area = polygon.area
|
||||
if area >= min_area * np.prod(image.shape[:2]) and area <= max_area * np.prod(image.shape[:2]) and hierarchy[0][jv][3] == -1: # and hierarchy[0][jv][3]==-1 :
|
||||
found_polygons_early.append(np.array([[point] for point in polygon.exterior.coords], dtype=np.uint))
|
||||
jv += 1
|
||||
return found_polygons_early
|
||||
|
||||
def filter_contours_area_of_image_tables(image, contours, hierarchy, max_area, min_area):
|
||||
found_polygons_early = list()
|
||||
|
||||
jv = 0
|
||||
for c in contours:
|
||||
if len(c) < 3: # A polygon cannot have less than 3 points
|
||||
continue
|
||||
|
||||
polygon = geometry.Polygon([point[0] for point in c])
|
||||
# area = cv2.contourArea(c)
|
||||
area = polygon.area
|
||||
##print(np.prod(thresh.shape[:2]))
|
||||
# Check that polygon has area greater than minimal area
|
||||
# print(hierarchy[0][jv][3],hierarchy )
|
||||
if area >= min_area * np.prod(image.shape[:2]) and area <= max_area * np.prod(image.shape[:2]): # and hierarchy[0][jv][3]==-1 :
|
||||
# print(c[0][0][1])
|
||||
found_polygons_early.append(np.array([[point] for point in polygon.exterior.coords], dtype=np.int32))
|
||||
jv += 1
|
||||
return found_polygons_early
|
||||
|
||||
def find_new_features_of_contours(contours_main):
|
||||
|
||||
areas_main = np.array([cv2.contourArea(contours_main[j]) for j in range(len(contours_main))])
|
||||
M_main = [cv2.moments(contours_main[j]) for j in range(len(contours_main))]
|
||||
cx_main = [(M_main[j]["m10"] / (M_main[j]["m00"] + 1e-32)) for j in range(len(M_main))]
|
||||
cy_main = [(M_main[j]["m01"] / (M_main[j]["m00"] + 1e-32)) for j in range(len(M_main))]
|
||||
try:
|
||||
x_min_main = np.array([np.min(contours_main[j][:, 0, 0]) for j in range(len(contours_main))])
|
||||
|
||||
argmin_x_main = np.array([np.argmin(contours_main[j][:, 0, 0]) for j in range(len(contours_main))])
|
||||
|
||||
x_min_from_argmin = np.array([contours_main[j][argmin_x_main[j], 0, 0] for j in range(len(contours_main))])
|
||||
y_corr_x_min_from_argmin = np.array([contours_main[j][argmin_x_main[j], 0, 1] for j in range(len(contours_main))])
|
||||
|
||||
x_max_main = np.array([np.max(contours_main[j][:, 0, 0]) for j in range(len(contours_main))])
|
||||
|
||||
y_min_main = np.array([np.min(contours_main[j][:, 0, 1]) for j in range(len(contours_main))])
|
||||
y_max_main = np.array([np.max(contours_main[j][:, 0, 1]) for j in range(len(contours_main))])
|
||||
except:
|
||||
x_min_main = np.array([np.min(contours_main[j][:, 0]) for j in range(len(contours_main))])
|
||||
|
||||
argmin_x_main = np.array([np.argmin(contours_main[j][:, 0]) for j in range(len(contours_main))])
|
||||
|
||||
x_min_from_argmin = np.array([contours_main[j][argmin_x_main[j], 0] for j in range(len(contours_main))])
|
||||
y_corr_x_min_from_argmin = np.array([contours_main[j][argmin_x_main[j], 1] for j in range(len(contours_main))])
|
||||
|
||||
x_max_main = np.array([np.max(contours_main[j][:, 0]) for j in range(len(contours_main))])
|
||||
|
||||
y_min_main = np.array([np.min(contours_main[j][:, 1]) for j in range(len(contours_main))])
|
||||
y_max_main = np.array([np.max(contours_main[j][:, 1]) for j in range(len(contours_main))])
|
||||
|
||||
# dis_x=np.abs(x_max_main-x_min_main)
|
||||
|
||||
return cx_main, cy_main, x_min_main, x_max_main, y_min_main, y_max_main, y_corr_x_min_from_argmin
|
||||
|
||||
def return_parent_contours(contours, hierarchy):
|
||||
contours_parent = [contours[i] for i in range(len(contours)) if hierarchy[0][i][3] == -1]
|
||||
return contours_parent
|
||||
|
||||
def return_contours_of_interested_region(region_pre_p, pixel, min_area=0.0002):
|
||||
|
||||
# pixels of images are identified by 5
|
||||
if len(region_pre_p.shape) == 3:
|
||||
cnts_images = (region_pre_p[:, :, 0] == pixel) * 1
|
||||
else:
|
||||
cnts_images = (region_pre_p[:, :] == pixel) * 1
|
||||
cnts_images = cnts_images.astype(np.uint8)
|
||||
cnts_images = np.repeat(cnts_images[:, :, np.newaxis], 3, axis=2)
|
||||
imgray = cv2.cvtColor(cnts_images, cv2.COLOR_BGR2GRAY)
|
||||
ret, thresh = cv2.threshold(imgray, 0, 255, 0)
|
||||
|
||||
contours_imgs, hierarchy = cv2.findContours(thresh, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
|
||||
|
||||
contours_imgs = return_parent_contours(contours_imgs, hierarchy)
|
||||
contours_imgs = filter_contours_area_of_image_tables(thresh, contours_imgs, hierarchy, max_area=1, min_area=min_area)
|
||||
|
||||
return contours_imgs
|
||||
|
||||
def get_textregion_contours_in_org_image(cnts, img, slope_first):
|
||||
|
||||
cnts_org = []
|
||||
# print(cnts,'cnts')
|
||||
for i in range(len(cnts)):
|
||||
img_copy = np.zeros(img.shape)
|
||||
img_copy = cv2.fillPoly(img_copy, pts=[cnts[i]], color=(1, 1, 1))
|
||||
|
||||
# plt.imshow(img_copy)
|
||||
# plt.show()
|
||||
|
||||
# print(img.shape,'img')
|
||||
img_copy = rotation_image_new(img_copy, -slope_first)
|
||||
##print(img_copy.shape,'img_copy')
|
||||
# plt.imshow(img_copy)
|
||||
# plt.show()
|
||||
|
||||
img_copy = img_copy.astype(np.uint8)
|
||||
imgray = cv2.cvtColor(img_copy, cv2.COLOR_BGR2GRAY)
|
||||
ret, thresh = cv2.threshold(imgray, 0, 255, 0)
|
||||
|
||||
cont_int, _ = cv2.findContours(thresh, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
|
||||
|
||||
cont_int[0][:, 0, 0] = cont_int[0][:, 0, 0] + np.abs(img_copy.shape[1] - img.shape[1])
|
||||
cont_int[0][:, 0, 1] = cont_int[0][:, 0, 1] + np.abs(img_copy.shape[0] - img.shape[0])
|
||||
# print(np.shape(cont_int[0]))
|
||||
cnts_org.append(cont_int[0])
|
||||
|
||||
# print(cnts_org,'cnts_org')
|
||||
|
||||
# sys.exit()
|
||||
# self.y_shift = np.abs(img_copy.shape[0] - img.shape[0])
|
||||
# self.x_shift = np.abs(img_copy.shape[1] - img.shape[1])
|
||||
return cnts_org
|
||||
|
||||
def return_contours_of_interested_textline(region_pre_p, pixel):
|
||||
|
||||
# pixels of images are identified by 5
|
||||
if len(region_pre_p.shape) == 3:
|
||||
cnts_images = (region_pre_p[:, :, 0] == pixel) * 1
|
||||
else:
|
||||
cnts_images = (region_pre_p[:, :] == pixel) * 1
|
||||
cnts_images = cnts_images.astype(np.uint8)
|
||||
cnts_images = np.repeat(cnts_images[:, :, np.newaxis], 3, axis=2)
|
||||
imgray = cv2.cvtColor(cnts_images, cv2.COLOR_BGR2GRAY)
|
||||
ret, thresh = cv2.threshold(imgray, 0, 255, 0)
|
||||
contours_imgs, hierarchy = cv2.findContours(thresh, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
|
||||
|
||||
contours_imgs = return_parent_contours(contours_imgs, hierarchy)
|
||||
contours_imgs = filter_contours_area_of_image_tables(thresh, contours_imgs, hierarchy, max_area=1, min_area=0.000000003)
|
||||
return contours_imgs
|
||||
|
||||
def return_contours_of_image(image):
|
||||
|
||||
if len(image.shape) == 2:
|
||||
image = np.repeat(image[:, :, np.newaxis], 3, axis=2)
|
||||
image = image.astype(np.uint8)
|
||||
else:
|
||||
image = image.astype(np.uint8)
|
||||
imgray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
|
||||
ret, thresh = cv2.threshold(imgray, 0, 255, 0)
|
||||
contours, hierarchy = cv2.findContours(thresh, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
|
||||
return contours, hierarchy
|
||||
|
||||
def return_contours_of_interested_region_by_min_size(region_pre_p, pixel, min_size=0.00003):
|
||||
|
||||
# pixels of images are identified by 5
|
||||
if len(region_pre_p.shape) == 3:
|
||||
cnts_images = (region_pre_p[:, :, 0] == pixel) * 1
|
||||
else:
|
||||
cnts_images = (region_pre_p[:, :] == pixel) * 1
|
||||
cnts_images = cnts_images.astype(np.uint8)
|
||||
cnts_images = np.repeat(cnts_images[:, :, np.newaxis], 3, axis=2)
|
||||
imgray = cv2.cvtColor(cnts_images, cv2.COLOR_BGR2GRAY)
|
||||
ret, thresh = cv2.threshold(imgray, 0, 255, 0)
|
||||
|
||||
contours_imgs, hierarchy = cv2.findContours(thresh, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
|
||||
|
||||
contours_imgs = return_parent_contours(contours_imgs, hierarchy)
|
||||
contours_imgs = filter_contours_area_of_image_tables(thresh, contours_imgs, hierarchy, max_area=1, min_area=min_size)
|
||||
|
||||
return contours_imgs
|
||||
|
||||
def return_contours_of_interested_region_by_size(region_pre_p, pixel, min_area, max_area):
|
||||
|
||||
# pixels of images are identified by 5
|
||||
if len(region_pre_p.shape) == 3:
|
||||
cnts_images = (region_pre_p[:, :, 0] == pixel) * 1
|
||||
else:
|
||||
cnts_images = (region_pre_p[:, :] == pixel) * 1
|
||||
cnts_images = cnts_images.astype(np.uint8)
|
||||
cnts_images = np.repeat(cnts_images[:, :, np.newaxis], 3, axis=2)
|
||||
imgray = cv2.cvtColor(cnts_images, cv2.COLOR_BGR2GRAY)
|
||||
ret, thresh = cv2.threshold(imgray, 0, 255, 0)
|
||||
contours_imgs, hierarchy = cv2.findContours(thresh, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
|
||||
|
||||
contours_imgs = return_parent_contours(contours_imgs, hierarchy)
|
||||
contours_imgs = filter_contours_area_of_image_tables(thresh, contours_imgs, hierarchy, max_area=max_area, min_area=min_area)
|
||||
|
||||
img_ret = np.zeros((region_pre_p.shape[0], region_pre_p.shape[1], 3))
|
||||
img_ret = cv2.fillPoly(img_ret, pts=contours_imgs, color=(1, 1, 1))
|
||||
return img_ret[:, :, 0]
|
||||
|
||||
|
|
@ -3,17 +3,17 @@ from collections import Counter
|
|||
REGION_ID_TEMPLATE = 'region_%04d'
|
||||
LINE_ID_TEMPLATE = 'region_%04d_line_%04d'
|
||||
|
||||
class EynollahIdCounter:
|
||||
class EynollahIdCounter():
|
||||
|
||||
def __init__(self, region_idx=0, line_idx=0):
|
||||
self._counter = Counter()
|
||||
self._initial_region_idx = region_idx
|
||||
self._initial_line_idx = line_idx
|
||||
self._inital_region_idx = region_idx
|
||||
self._inital_line_idx = line_idx
|
||||
self.reset()
|
||||
|
||||
def reset(self):
|
||||
self.set('region', self._initial_region_idx)
|
||||
self.set('line', self._initial_line_idx)
|
||||
self.set('region', self._inital_region_idx)
|
||||
self.set('line', self._inital_line_idx)
|
||||
|
||||
def inc(self, name, val=1):
|
||||
self._counter.update({name: val})
|
||||
|
|
@ -1,11 +1,9 @@
|
|||
import numpy as np
|
||||
import cv2
|
||||
from .contour import (
|
||||
find_center_of_contours,
|
||||
find_new_features_of_contours,
|
||||
return_contours_of_image,
|
||||
return_parent_contours,
|
||||
return_contours_of_interested_region,
|
||||
)
|
||||
|
||||
def adhere_drop_capital_region_into_corresponding_textline(
|
||||
|
|
@ -15,15 +13,15 @@ def adhere_drop_capital_region_into_corresponding_textline(
|
|||
contours_only_text_parent_h,
|
||||
all_box_coord,
|
||||
all_box_coord_h,
|
||||
all_found_textline_polygons,
|
||||
all_found_textline_polygons_h,
|
||||
all_found_texline_polygons,
|
||||
all_found_texline_polygons_h,
|
||||
kernel=None,
|
||||
curved_line=False,
|
||||
):
|
||||
# print(np.shape(all_found_textline_polygons),np.shape(all_found_textline_polygons[3]),'all_found_textline_polygonsshape')
|
||||
# print(all_found_textline_polygons[3])
|
||||
cx_m, cy_m = find_center_of_contours(contours_only_text_parent)
|
||||
cx_h, cy_h = find_center_of_contours(contours_only_text_parent_h)
|
||||
# print(np.shape(all_found_texline_polygons),np.shape(all_found_texline_polygons[3]),'all_found_texline_polygonsshape')
|
||||
# print(all_found_texline_polygons[3])
|
||||
cx_m, cy_m, _, _, _, _, _ = find_new_features_of_contours(contours_only_text_parent)
|
||||
cx_h, cy_h, _, _, _, _, _ = find_new_features_of_contours(contours_only_text_parent_h)
|
||||
cx_d, cy_d, _, _, y_min_d, y_max_d, _ = find_new_features_of_contours(polygons_of_drop_capitals)
|
||||
|
||||
img_con_all = np.zeros((text_regions_p.shape[0], text_regions_p.shape[1], 3))
|
||||
|
|
@ -89,9 +87,9 @@ def adhere_drop_capital_region_into_corresponding_textline(
|
|||
region_final = region_with_intersected_drop[np.argmax(sum_pixels_of_intersection)] - 1
|
||||
|
||||
# print(region_final,'region_final')
|
||||
# cx_t, cy_t = find_center_of_contours(all_found_textline_polygons[int(region_final)])
|
||||
# cx_t,cy_t ,_, _, _ ,_,_= find_new_features_of_contours(all_found_texline_polygons[int(region_final)])
|
||||
try:
|
||||
cx_t, cy_t = find_center_of_contours(all_found_textline_polygons[int(region_final)])
|
||||
cx_t, cy_t, _, _, _, _, _ = find_new_features_of_contours(all_found_texline_polygons[int(region_final)])
|
||||
# print(all_box_coord[j_cont])
|
||||
# print(cx_t)
|
||||
# print(cy_t)
|
||||
|
|
@ -107,26 +105,21 @@ def adhere_drop_capital_region_into_corresponding_textline(
|
|||
arg_min = np.argmin(np.abs(y_lines - y_min_d[i_drop]))
|
||||
# print(arg_min)
|
||||
|
||||
cnt_nearest = np.copy(all_found_textline_polygons[int(region_final)][arg_min])
|
||||
cnt_nearest[:, 0, 0] = all_found_textline_polygons[int(region_final)][arg_min][:, 0, 0] # +all_box_coord[int(region_final)][2]
|
||||
cnt_nearest[:, 0, 1] = all_found_textline_polygons[int(region_final)][arg_min][:, 0, 1] # +all_box_coord[int(region_final)][0]
|
||||
cnt_nearest = np.copy(all_found_texline_polygons[int(region_final)][arg_min])
|
||||
cnt_nearest[:, 0, 0] = all_found_texline_polygons[int(region_final)][arg_min][:, 0, 0] # +all_box_coord[int(region_final)][2]
|
||||
cnt_nearest[:, 0, 1] = all_found_texline_polygons[int(region_final)][arg_min][:, 0, 1] # +all_box_coord[int(region_final)][0]
|
||||
|
||||
img_textlines = np.zeros((text_regions_p.shape[0], text_regions_p.shape[1], 3))
|
||||
img_textlines = cv2.fillPoly(img_textlines, pts=[cnt_nearest], color=(255, 255, 255))
|
||||
img_textlines = cv2.fillPoly(img_textlines, pts=[polygons_of_drop_capitals[i_drop]], color=(255, 255, 255))
|
||||
|
||||
img_textlines = img_textlines.astype(np.uint8)
|
||||
imgray = cv2.cvtColor(img_textlines, cv2.COLOR_BGR2GRAY)
|
||||
ret, thresh = cv2.threshold(imgray, 0, 255, 0)
|
||||
|
||||
contours_combined = return_contours_of_interested_region(img_textlines, 255, 0)
|
||||
|
||||
#plt.imshow(img_textlines)
|
||||
#plt.show()
|
||||
|
||||
#imgray = cv2.cvtColor(img_textlines, cv2.COLOR_BGR2GRAY)
|
||||
#ret, thresh = cv2.threshold(imgray, 0, 255, 0)
|
||||
|
||||
#contours_combined, hierarchy = cv2.findContours(thresh, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
|
||||
contours_combined, hierarchy = cv2.findContours(thresh, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
|
||||
|
||||
# print(len(contours_combined),'len textlines mixed')
|
||||
areas_cnt_text = np.array([cv2.contourArea(contours_combined[j]) for j in range(len(contours_combined))])
|
||||
|
||||
contours_biggest = contours_combined[np.argmax(areas_cnt_text)]
|
||||
|
|
@ -138,12 +131,7 @@ def adhere_drop_capital_region_into_corresponding_textline(
|
|||
|
||||
# contours_biggest=contours_biggest.reshape(np.shape(contours_biggest)[0],np.shape(contours_biggest)[2])
|
||||
|
||||
if len(contours_combined)==1:
|
||||
all_found_textline_polygons[int(region_final)][arg_min] = contours_biggest
|
||||
elif len(contours_combined)==2:
|
||||
all_found_textline_polygons[int(region_final)].insert(arg_min, polygons_of_drop_capitals[i_drop] )
|
||||
else:
|
||||
pass
|
||||
all_found_texline_polygons[int(region_final)][arg_min] = contours_biggest
|
||||
|
||||
except:
|
||||
# print('gordun1')
|
||||
|
|
@ -151,11 +139,11 @@ def adhere_drop_capital_region_into_corresponding_textline(
|
|||
elif len(region_with_intersected_drop) == 1:
|
||||
region_final = region_with_intersected_drop[0] - 1
|
||||
|
||||
# areas_main=np.array([cv2.contourArea(all_found_textline_polygons[int(region_final)][0][j] ) for j in range(len(all_found_textline_polygons[int(region_final)]))])
|
||||
# areas_main=np.array([cv2.contourArea(all_found_texline_polygons[int(region_final)][0][j] ) for j in range(len(all_found_texline_polygons[int(region_final)]))])
|
||||
|
||||
# cx_t, cy_t = find_center_of_contours(all_found_textline_polygons[int(region_final)])
|
||||
try:
|
||||
cx_t, cy_t = find_center_of_contours(all_found_textline_polygons[int(region_final)])
|
||||
# cx_t,cy_t ,_, _, _ ,_,_= find_new_features_of_contours(all_found_texline_polygons[int(region_final)])
|
||||
|
||||
cx_t, cy_t, _, _, _, _, _ = find_new_features_of_contours(all_found_texline_polygons[int(region_final)])
|
||||
# print(all_box_coord[j_cont])
|
||||
# print(cx_t)
|
||||
# print(cy_t)
|
||||
|
|
@ -169,9 +157,9 @@ def adhere_drop_capital_region_into_corresponding_textline(
|
|||
arg_min = np.argmin(np.abs(y_lines - y_min_d[i_drop]))
|
||||
# print(arg_min)
|
||||
|
||||
cnt_nearest = np.copy(all_found_textline_polygons[int(region_final)][arg_min])
|
||||
cnt_nearest[:, 0, 0] = all_found_textline_polygons[int(region_final)][arg_min][:, 0, 0] # +all_box_coord[int(region_final)][2]
|
||||
cnt_nearest[:, 0, 1] = all_found_textline_polygons[int(region_final)][arg_min][:, 0, 1] # +all_box_coord[int(region_final)][0]
|
||||
cnt_nearest = np.copy(all_found_texline_polygons[int(region_final)][arg_min])
|
||||
cnt_nearest[:, 0, 0] = all_found_texline_polygons[int(region_final)][arg_min][:, 0, 0] # +all_box_coord[int(region_final)][2]
|
||||
cnt_nearest[:, 0, 1] = all_found_texline_polygons[int(region_final)][arg_min][:, 0, 1] # +all_box_coord[int(region_final)][0]
|
||||
|
||||
img_textlines = np.zeros((text_regions_p.shape[0], text_regions_p.shape[1], 3))
|
||||
img_textlines = cv2.fillPoly(img_textlines, pts=[cnt_nearest], color=(255, 255, 255))
|
||||
|
|
@ -179,13 +167,14 @@ def adhere_drop_capital_region_into_corresponding_textline(
|
|||
|
||||
img_textlines = img_textlines.astype(np.uint8)
|
||||
|
||||
# plt.imshow(img_textlines)
|
||||
# plt.show()
|
||||
imgray = cv2.cvtColor(img_textlines, cv2.COLOR_BGR2GRAY)
|
||||
ret, thresh = cv2.threshold(imgray, 0, 255, 0)
|
||||
|
||||
contours_combined = return_contours_of_interested_region(img_textlines, 255, 0)
|
||||
##imgray = cv2.cvtColor(img_textlines, cv2.COLOR_BGR2GRAY)
|
||||
##ret, thresh = cv2.threshold(imgray, 0, 255, 0)
|
||||
|
||||
##contours_combined, hierarchy = cv2.findContours(thresh, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
|
||||
contours_combined, hierarchy = cv2.findContours(thresh, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
|
||||
|
||||
# print(len(contours_combined),'len textlines mixed')
|
||||
areas_cnt_text = np.array([cv2.contourArea(contours_combined[j]) for j in range(len(contours_combined))])
|
||||
|
||||
contours_biggest = contours_combined[np.argmax(areas_cnt_text)]
|
||||
|
|
@ -195,20 +184,14 @@ def adhere_drop_capital_region_into_corresponding_textline(
|
|||
# contours_biggest[:,0,0]=contours_biggest[:,0,0]#-all_box_coord[int(region_final)][2]
|
||||
# contours_biggest[:,0,1]=contours_biggest[:,0,1]#-all_box_coord[int(region_final)][0]
|
||||
# print(np.shape(contours_biggest),'contours_biggest')
|
||||
# print(np.shape(all_found_textline_polygons[int(region_final)][arg_min]))
|
||||
# print(np.shape(all_found_texline_polygons[int(region_final)][arg_min]))
|
||||
##contours_biggest=contours_biggest.reshape(np.shape(contours_biggest)[0],np.shape(contours_biggest)[2])
|
||||
if len(contours_combined)==1:
|
||||
all_found_textline_polygons[int(region_final)][arg_min] = contours_biggest
|
||||
elif len(contours_combined)==2:
|
||||
all_found_textline_polygons[int(region_final)].insert(arg_min, polygons_of_drop_capitals[i_drop] )
|
||||
else:
|
||||
pass
|
||||
except:
|
||||
pass
|
||||
all_found_texline_polygons[int(region_final)][arg_min] = contours_biggest
|
||||
|
||||
# print(cx_t,'print')
|
||||
try:
|
||||
# print(all_found_textline_polygons[j_cont][0])
|
||||
cx_t, cy_t = find_center_of_contours(all_found_textline_polygons[int(region_final)])
|
||||
# print(all_found_texline_polygons[j_cont][0])
|
||||
cx_t, cy_t, _, _, _, _, _ = find_new_features_of_contours(all_found_texline_polygons[int(region_final)])
|
||||
# print(all_box_coord[j_cont])
|
||||
# print(cx_t)
|
||||
# print(cy_t)
|
||||
|
|
@ -222,20 +205,19 @@ def adhere_drop_capital_region_into_corresponding_textline(
|
|||
arg_min = np.argmin(np.abs(y_lines - y_min_d[i_drop]))
|
||||
# print(arg_min)
|
||||
|
||||
cnt_nearest = np.copy(all_found_textline_polygons[int(region_final)][arg_min])
|
||||
cnt_nearest[:, 0, 0] = all_found_textline_polygons[int(region_final)][arg_min][:, 0, 0] # +all_box_coord[int(region_final)][2]
|
||||
cnt_nearest[:, 0, 1] = all_found_textline_polygons[int(region_final)][arg_min][:, 0, 1] # +all_box_coord[int(region_final)][0]
|
||||
cnt_nearest = np.copy(all_found_texline_polygons[int(region_final)][arg_min])
|
||||
cnt_nearest[:, 0, 0] = all_found_texline_polygons[int(region_final)][arg_min][:, 0, 0] # +all_box_coord[int(region_final)][2]
|
||||
cnt_nearest[:, 0, 1] = all_found_texline_polygons[int(region_final)][arg_min][:, 0, 1] # +all_box_coord[int(region_final)][0]
|
||||
|
||||
img_textlines = np.zeros((text_regions_p.shape[0], text_regions_p.shape[1], 3))
|
||||
img_textlines = cv2.fillPoly(img_textlines, pts=[cnt_nearest], color=(255, 255, 255))
|
||||
img_textlines = cv2.fillPoly(img_textlines, pts=[polygons_of_drop_capitals[i_drop]], color=(255, 255, 255))
|
||||
|
||||
img_textlines = img_textlines.astype(np.uint8)
|
||||
contours_combined = return_contours_of_interested_region(img_textlines, 255, 0)
|
||||
#imgray = cv2.cvtColor(img_textlines, cv2.COLOR_BGR2GRAY)
|
||||
#ret, thresh = cv2.threshold(imgray, 0, 255, 0)
|
||||
imgray = cv2.cvtColor(img_textlines, cv2.COLOR_BGR2GRAY)
|
||||
ret, thresh = cv2.threshold(imgray, 0, 255, 0)
|
||||
|
||||
#contours_combined, hierarchy = cv2.findContours(thresh, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
|
||||
contours_combined, hierarchy = cv2.findContours(thresh, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
|
||||
|
||||
# print(len(contours_combined),'len textlines mixed')
|
||||
areas_cnt_text = np.array([cv2.contourArea(contours_combined[j]) for j in range(len(contours_combined))])
|
||||
|
|
@ -248,20 +230,15 @@ def adhere_drop_capital_region_into_corresponding_textline(
|
|||
contours_biggest[:, 0, 1] = contours_biggest[:, 0, 1] # -all_box_coord[int(region_final)][0]
|
||||
|
||||
##contours_biggest=contours_biggest.reshape(np.shape(contours_biggest)[0],np.shape(contours_biggest)[2])
|
||||
if len(contours_combined)==1:
|
||||
all_found_textline_polygons[int(region_final)][arg_min] = contours_biggest
|
||||
elif len(contours_combined)==2:
|
||||
all_found_textline_polygons[int(region_final)].insert(arg_min, polygons_of_drop_capitals[i_drop] )
|
||||
else:
|
||||
pass
|
||||
# all_found_textline_polygons[int(region_final)][arg_min]=contours_biggest
|
||||
all_found_texline_polygons[int(region_final)][arg_min] = contours_biggest
|
||||
# all_found_texline_polygons[int(region_final)][arg_min]=contours_biggest
|
||||
|
||||
except:
|
||||
pass
|
||||
else:
|
||||
pass
|
||||
|
||||
##cx_t, cy_t = find_center_of_contours(all_found_textline_polygons[int(region_final)])
|
||||
##cx_t,cy_t ,_, _, _ ,_,_= find_new_features_of_contours(all_found_texline_polygons[int(region_final)])
|
||||
###print(all_box_coord[j_cont])
|
||||
###print(cx_t)
|
||||
###print(cy_t)
|
||||
|
|
@ -275,9 +252,9 @@ def adhere_drop_capital_region_into_corresponding_textline(
|
|||
##arg_min=np.argmin(np.abs(y_lines-y_min_d[i_drop]) )
|
||||
###print(arg_min)
|
||||
|
||||
##cnt_nearest=np.copy(all_found_textline_polygons[int(region_final)][arg_min])
|
||||
##cnt_nearest[:,0,0]=all_found_textline_polygons[int(region_final)][arg_min][:,0,0]#+all_box_coord[int(region_final)][2]
|
||||
##cnt_nearest[:,0,1]=all_found_textline_polygons[int(region_final)][arg_min][:,0,1]#+all_box_coord[int(region_final)][0]
|
||||
##cnt_nearest=np.copy(all_found_texline_polygons[int(region_final)][arg_min])
|
||||
##cnt_nearest[:,0,0]=all_found_texline_polygons[int(region_final)][arg_min][:,0,0]#+all_box_coord[int(region_final)][2]
|
||||
##cnt_nearest[:,0,1]=all_found_texline_polygons[int(region_final)][arg_min][:,0,1]#+all_box_coord[int(region_final)][0]
|
||||
|
||||
##img_textlines=np.zeros((text_regions_p.shape[0],text_regions_p.shape[1],3))
|
||||
##img_textlines=cv2.fillPoly(img_textlines,pts=[cnt_nearest],color=(255,255,255))
|
||||
|
|
@ -303,7 +280,7 @@ def adhere_drop_capital_region_into_corresponding_textline(
|
|||
##contours_biggest[:,0,1]=contours_biggest[:,0,1]#-all_box_coord[int(region_final)][0]
|
||||
|
||||
##contours_biggest=contours_biggest.reshape(np.shape(contours_biggest)[0],np.shape(contours_biggest)[2])
|
||||
##all_found_textline_polygons[int(region_final)][arg_min]=contours_biggest
|
||||
##all_found_texline_polygons[int(region_final)][arg_min]=contours_biggest
|
||||
|
||||
else:
|
||||
if len(region_with_intersected_drop) > 1:
|
||||
|
|
@ -315,9 +292,9 @@ def adhere_drop_capital_region_into_corresponding_textline(
|
|||
region_final = region_with_intersected_drop[np.argmax(sum_pixels_of_intersection)] - 1
|
||||
|
||||
# print(region_final,'region_final')
|
||||
# cx_t, cy_t = find_center_of_contours(all_found_textline_polygons[int(region_final)])
|
||||
# cx_t,cy_t ,_, _, _ ,_,_= find_new_features_of_contours(all_found_texline_polygons[int(region_final)])
|
||||
try:
|
||||
cx_t, cy_t = find_center_of_contours(all_found_textline_polygons[int(region_final)])
|
||||
cx_t, cy_t, _, _, _, _, _ = find_new_features_of_contours(all_found_texline_polygons[int(region_final)])
|
||||
# print(all_box_coord[j_cont])
|
||||
# print(cx_t)
|
||||
# print(cy_t)
|
||||
|
|
@ -333,21 +310,19 @@ def adhere_drop_capital_region_into_corresponding_textline(
|
|||
arg_min = np.argmin(np.abs(y_lines - y_min_d[i_drop]))
|
||||
# print(arg_min)
|
||||
|
||||
cnt_nearest = np.copy(all_found_textline_polygons[int(region_final)][arg_min])
|
||||
cnt_nearest[:, 0] = all_found_textline_polygons[int(region_final)][arg_min][:, 0] + all_box_coord[int(region_final)][2]
|
||||
cnt_nearest[:, 1] = all_found_textline_polygons[int(region_final)][arg_min][:, 1] + all_box_coord[int(region_final)][0]
|
||||
cnt_nearest = np.copy(all_found_texline_polygons[int(region_final)][arg_min])
|
||||
cnt_nearest[:, 0] = all_found_texline_polygons[int(region_final)][arg_min][:, 0] + all_box_coord[int(region_final)][2]
|
||||
cnt_nearest[:, 1] = all_found_texline_polygons[int(region_final)][arg_min][:, 1] + all_box_coord[int(region_final)][0]
|
||||
|
||||
img_textlines = np.zeros((text_regions_p.shape[0], text_regions_p.shape[1], 3))
|
||||
img_textlines = cv2.fillPoly(img_textlines, pts=[cnt_nearest], color=(255, 255, 255))
|
||||
img_textlines = cv2.fillPoly(img_textlines, pts=[polygons_of_drop_capitals[i_drop]], color=(255, 255, 255))
|
||||
|
||||
img_textlines = img_textlines.astype(np.uint8)
|
||||
contours_combined = return_contours_of_interested_region(img_textlines, 255, 0)
|
||||
imgray = cv2.cvtColor(img_textlines, cv2.COLOR_BGR2GRAY)
|
||||
ret, thresh = cv2.threshold(imgray, 0, 255, 0)
|
||||
|
||||
#imgray = cv2.cvtColor(img_textlines, cv2.COLOR_BGR2GRAY)
|
||||
#ret, thresh = cv2.threshold(imgray, 0, 255, 0)
|
||||
|
||||
#contours_combined, hierarchy = cv2.findContours(thresh, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
|
||||
contours_combined, hierarchy = cv2.findContours(thresh, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
|
||||
|
||||
# print(len(contours_combined),'len textlines mixed')
|
||||
areas_cnt_text = np.array([cv2.contourArea(contours_combined[j]) for j in range(len(contours_combined))])
|
||||
|
|
@ -360,12 +335,8 @@ def adhere_drop_capital_region_into_corresponding_textline(
|
|||
contours_biggest[:, 0, 1] = contours_biggest[:, 0, 1] - all_box_coord[int(region_final)][0]
|
||||
|
||||
contours_biggest = contours_biggest.reshape(np.shape(contours_biggest)[0], np.shape(contours_biggest)[2])
|
||||
if len(contours_combined)==1:
|
||||
all_found_textline_polygons[int(region_final)][arg_min] = contours_biggest
|
||||
elif len(contours_combined)==2:
|
||||
all_found_textline_polygons[int(region_final)].insert(arg_min, polygons_of_drop_capitals[i_drop] )
|
||||
else:
|
||||
pass
|
||||
|
||||
all_found_texline_polygons[int(region_final)][arg_min] = contours_biggest
|
||||
|
||||
except:
|
||||
# print('gordun1')
|
||||
|
|
@ -373,14 +344,14 @@ def adhere_drop_capital_region_into_corresponding_textline(
|
|||
elif len(region_with_intersected_drop) == 1:
|
||||
region_final = region_with_intersected_drop[0] - 1
|
||||
|
||||
# areas_main=np.array([cv2.contourArea(all_found_textline_polygons[int(region_final)][0][j] ) for j in range(len(all_found_textline_polygons[int(region_final)]))])
|
||||
# areas_main=np.array([cv2.contourArea(all_found_texline_polygons[int(region_final)][0][j] ) for j in range(len(all_found_texline_polygons[int(region_final)]))])
|
||||
|
||||
# cx_t, cy_t = find_center_of_contours(all_found_textline_polygons[int(region_final)])
|
||||
# cx_t,cy_t ,_, _, _ ,_,_= find_new_features_of_contours(all_found_texline_polygons[int(region_final)])
|
||||
|
||||
# print(cx_t,'print')
|
||||
try:
|
||||
# print(all_found_textline_polygons[j_cont][0])
|
||||
cx_t, cy_t = find_center_of_contours(all_found_textline_polygons[int(region_final)])
|
||||
# print(all_found_texline_polygons[j_cont][0])
|
||||
cx_t, cy_t, _, _, _, _, _ = find_new_features_of_contours(all_found_texline_polygons[int(region_final)])
|
||||
# print(all_box_coord[j_cont])
|
||||
# print(cx_t)
|
||||
# print(cy_t)
|
||||
|
|
@ -394,21 +365,19 @@ def adhere_drop_capital_region_into_corresponding_textline(
|
|||
arg_min = np.argmin(np.abs(y_lines - y_min_d[i_drop]))
|
||||
# print(arg_min)
|
||||
|
||||
cnt_nearest = np.copy(all_found_textline_polygons[int(region_final)][arg_min])
|
||||
cnt_nearest[:, 0] = all_found_textline_polygons[int(region_final)][arg_min][:, 0] + all_box_coord[int(region_final)][2]
|
||||
cnt_nearest[:, 1] = all_found_textline_polygons[int(region_final)][arg_min][:, 1] + all_box_coord[int(region_final)][0]
|
||||
cnt_nearest = np.copy(all_found_texline_polygons[int(region_final)][arg_min])
|
||||
cnt_nearest[:, 0] = all_found_texline_polygons[int(region_final)][arg_min][:, 0] + all_box_coord[int(region_final)][2]
|
||||
cnt_nearest[:, 1] = all_found_texline_polygons[int(region_final)][arg_min][:, 1] + all_box_coord[int(region_final)][0]
|
||||
|
||||
img_textlines = np.zeros((text_regions_p.shape[0], text_regions_p.shape[1], 3))
|
||||
img_textlines = cv2.fillPoly(img_textlines, pts=[cnt_nearest], color=(255, 255, 255))
|
||||
img_textlines = cv2.fillPoly(img_textlines, pts=[polygons_of_drop_capitals[i_drop]], color=(255, 255, 255))
|
||||
|
||||
img_textlines = img_textlines.astype(np.uint8)
|
||||
contours_combined = return_contours_of_interested_region(img_textlines, 255, 0)
|
||||
imgray = cv2.cvtColor(img_textlines, cv2.COLOR_BGR2GRAY)
|
||||
ret, thresh = cv2.threshold(imgray, 0, 255, 0)
|
||||
|
||||
#imgray = cv2.cvtColor(img_textlines, cv2.COLOR_BGR2GRAY)
|
||||
#ret, thresh = cv2.threshold(imgray, 0, 255, 0)
|
||||
|
||||
#contours_combined, hierarchy = cv2.findContours(thresh, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
|
||||
contours_combined, hierarchy = cv2.findContours(thresh, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
|
||||
|
||||
# print(len(contours_combined),'len textlines mixed')
|
||||
areas_cnt_text = np.array([cv2.contourArea(contours_combined[j]) for j in range(len(contours_combined))])
|
||||
|
|
@ -421,13 +390,8 @@ def adhere_drop_capital_region_into_corresponding_textline(
|
|||
contours_biggest[:, 0, 1] = contours_biggest[:, 0, 1] - all_box_coord[int(region_final)][0]
|
||||
|
||||
contours_biggest = contours_biggest.reshape(np.shape(contours_biggest)[0], np.shape(contours_biggest)[2])
|
||||
if len(contours_combined)==1:
|
||||
all_found_textline_polygons[int(region_final)][arg_min] = contours_biggest
|
||||
elif len(contours_combined)==2:
|
||||
all_found_textline_polygons[int(region_final)].insert(arg_min, polygons_of_drop_capitals[i_drop] )
|
||||
else:
|
||||
pass
|
||||
# all_found_textline_polygons[int(region_final)][arg_min]=contours_biggest
|
||||
all_found_texline_polygons[int(region_final)][arg_min] = contours_biggest
|
||||
# all_found_texline_polygons[int(region_final)][arg_min]=contours_biggest
|
||||
|
||||
except:
|
||||
pass
|
||||
|
|
@ -452,8 +416,8 @@ def adhere_drop_capital_region_into_corresponding_textline(
|
|||
######plt.show()
|
||||
#####try:
|
||||
#####if len(contours_new_parent)==1:
|
||||
######print(all_found_textline_polygons[j_cont][0])
|
||||
#####cx_t, cy_t = find_center_of_contours(all_found_textline_polygons[j_cont])
|
||||
######print(all_found_texline_polygons[j_cont][0])
|
||||
#####cx_t,cy_t ,_, _, _ ,_,_= find_new_features_of_contours(all_found_texline_polygons[j_cont])
|
||||
######print(all_box_coord[j_cont])
|
||||
######print(cx_t)
|
||||
######print(cy_t)
|
||||
|
|
@ -466,9 +430,9 @@ def adhere_drop_capital_region_into_corresponding_textline(
|
|||
#####arg_min=np.argmin(np.abs(y_lines-y_min_d[i_drop]) )
|
||||
######print(arg_min)
|
||||
|
||||
#####cnt_nearest=np.copy(all_found_textline_polygons[j_cont][arg_min])
|
||||
#####cnt_nearest[:,0]=all_found_textline_polygons[j_cont][arg_min][:,0]+all_box_coord[j_cont][2]
|
||||
#####cnt_nearest[:,1]=all_found_textline_polygons[j_cont][arg_min][:,1]+all_box_coord[j_cont][0]
|
||||
#####cnt_nearest=np.copy(all_found_texline_polygons[j_cont][arg_min])
|
||||
#####cnt_nearest[:,0]=all_found_texline_polygons[j_cont][arg_min][:,0]+all_box_coord[j_cont][2]
|
||||
#####cnt_nearest[:,1]=all_found_texline_polygons[j_cont][arg_min][:,1]+all_box_coord[j_cont][0]
|
||||
|
||||
#####img_textlines=np.zeros((text_regions_p.shape[0],text_regions_p.shape[1],3))
|
||||
#####img_textlines=cv2.fillPoly(img_textlines,pts=[cnt_nearest],color=(255,255,255))
|
||||
|
|
@ -489,7 +453,7 @@ def adhere_drop_capital_region_into_corresponding_textline(
|
|||
#####contours_biggest[:,0,0]=contours_biggest[:,0,0]-all_box_coord[j_cont][2]
|
||||
#####contours_biggest[:,0,1]=contours_biggest[:,0,1]-all_box_coord[j_cont][0]
|
||||
|
||||
#####all_found_textline_polygons[j_cont][arg_min]=contours_biggest
|
||||
#####all_found_texline_polygons[j_cont][arg_min]=contours_biggest
|
||||
######print(contours_biggest)
|
||||
######plt.imshow(img_textlines[:,:,0])
|
||||
######plt.show()
|
||||
|
|
@ -497,11 +461,11 @@ def adhere_drop_capital_region_into_corresponding_textline(
|
|||
#####pass
|
||||
#####except:
|
||||
#####pass
|
||||
return all_found_textline_polygons
|
||||
return all_found_texline_polygons
|
||||
|
||||
def filter_small_drop_capitals_from_no_patch_layout(layout_no_patch, layout1):
|
||||
|
||||
drop_only = (layout_no_patch == 4) * 1
|
||||
drop_only = (layout_no_patch[:, :, 0] == 4) * 1
|
||||
contours_drop, hir_on_drop = return_contours_of_image(drop_only)
|
||||
contours_drop_parent = return_parent_contours(contours_drop, hir_on_drop)
|
||||
|
||||
|
|
@ -529,8 +493,9 @@ def filter_small_drop_capitals_from_no_patch_layout(layout_no_patch, layout1):
|
|||
if (((map_of_drop_contour_bb == 1) * 1).sum() / float(((map_of_drop_contour_bb == 5) * 1).sum()) * 100) >= 15:
|
||||
contours_drop_parent_final.append(contours_drop_parent[jj])
|
||||
|
||||
layout_no_patch[layout_no_patch == 4] = 0
|
||||
layout_no_patch = cv2.fillPoly(layout_no_patch, pts=contours_drop_parent_final, color=4)
|
||||
layout_no_patch[:, :, 0][layout_no_patch[:, :, 0] == 4] = 0
|
||||
|
||||
layout_no_patch = cv2.fillPoly(layout_no_patch, pts=contours_drop_parent_final, color=(4, 4, 4))
|
||||
|
||||
return layout_no_patch
|
||||
|
||||
252
qurator/eynollah/utils/marginals.py
Normal file
252
qurator/eynollah/utils/marginals.py
Normal file
|
|
@ -0,0 +1,252 @@
|
|||
import numpy as np
|
||||
import cv2
|
||||
from scipy.signal import find_peaks
|
||||
from scipy.ndimage import gaussian_filter1d
|
||||
|
||||
|
||||
from .contour import find_new_features_of_contours, return_contours_of_interested_region
|
||||
from .resize import resize_image
|
||||
from .rotate import rotate_image
|
||||
|
||||
def get_marginals(text_with_lines, text_regions, num_col, slope_deskew, kernel=None):
|
||||
mask_marginals=np.zeros((text_with_lines.shape[0],text_with_lines.shape[1]))
|
||||
mask_marginals=mask_marginals.astype(np.uint8)
|
||||
|
||||
|
||||
text_with_lines=text_with_lines.astype(np.uint8)
|
||||
##text_with_lines=cv2.erode(text_with_lines,self.kernel,iterations=3)
|
||||
|
||||
text_with_lines_eroded=cv2.erode(text_with_lines,kernel,iterations=5)
|
||||
|
||||
if text_with_lines.shape[0]<=1500:
|
||||
pass
|
||||
elif text_with_lines.shape[0]>1500 and text_with_lines.shape[0]<=1800:
|
||||
text_with_lines=resize_image(text_with_lines,int(text_with_lines.shape[0]*1.5),text_with_lines.shape[1])
|
||||
text_with_lines=cv2.erode(text_with_lines,kernel,iterations=5)
|
||||
text_with_lines=resize_image(text_with_lines,text_with_lines_eroded.shape[0],text_with_lines_eroded.shape[1])
|
||||
else:
|
||||
text_with_lines=resize_image(text_with_lines,int(text_with_lines.shape[0]*1.8),text_with_lines.shape[1])
|
||||
text_with_lines=cv2.erode(text_with_lines,kernel,iterations=7)
|
||||
text_with_lines=resize_image(text_with_lines,text_with_lines_eroded.shape[0],text_with_lines_eroded.shape[1])
|
||||
|
||||
|
||||
text_with_lines_y=text_with_lines.sum(axis=0)
|
||||
text_with_lines_y_eroded=text_with_lines_eroded.sum(axis=0)
|
||||
|
||||
thickness_along_y_percent=text_with_lines_y_eroded.max()/(float(text_with_lines.shape[0]))*100
|
||||
|
||||
#print(thickness_along_y_percent,'thickness_along_y_percent')
|
||||
|
||||
if thickness_along_y_percent<30:
|
||||
min_textline_thickness=8
|
||||
elif thickness_along_y_percent>=30 and thickness_along_y_percent<50:
|
||||
min_textline_thickness=20
|
||||
else:
|
||||
min_textline_thickness=40
|
||||
|
||||
|
||||
|
||||
if thickness_along_y_percent>=14:
|
||||
|
||||
text_with_lines_y_rev=-1*text_with_lines_y[:]
|
||||
#print(text_with_lines_y)
|
||||
#print(text_with_lines_y_rev)
|
||||
|
||||
|
||||
|
||||
|
||||
#plt.plot(text_with_lines_y)
|
||||
#plt.show()
|
||||
|
||||
|
||||
text_with_lines_y_rev=text_with_lines_y_rev-np.min(text_with_lines_y_rev)
|
||||
|
||||
#plt.plot(text_with_lines_y_rev)
|
||||
#plt.show()
|
||||
sigma_gaus=1
|
||||
region_sum_0= gaussian_filter1d(text_with_lines_y, sigma_gaus)
|
||||
|
||||
region_sum_0_rev=gaussian_filter1d(text_with_lines_y_rev, sigma_gaus)
|
||||
|
||||
#plt.plot(region_sum_0_rev)
|
||||
#plt.show()
|
||||
region_sum_0_updown=region_sum_0[len(region_sum_0)::-1]
|
||||
|
||||
first_nonzero=(next((i for i, x in enumerate(region_sum_0) if x), None))
|
||||
last_nonzero=(next((i for i, x in enumerate(region_sum_0_updown) if x), None))
|
||||
|
||||
|
||||
last_nonzero=len(region_sum_0)-last_nonzero
|
||||
|
||||
##img_sum_0_smooth_rev=-region_sum_0
|
||||
|
||||
|
||||
mid_point=(last_nonzero+first_nonzero)/2.
|
||||
|
||||
|
||||
one_third_right=(last_nonzero-mid_point)/3.0
|
||||
one_third_left=(mid_point-first_nonzero)/3.0
|
||||
|
||||
#img_sum_0_smooth_rev=img_sum_0_smooth_rev-np.min(img_sum_0_smooth_rev)
|
||||
|
||||
|
||||
|
||||
|
||||
peaks, _ = find_peaks(text_with_lines_y_rev, height=0)
|
||||
|
||||
|
||||
peaks=np.array(peaks)
|
||||
|
||||
|
||||
#print(region_sum_0[peaks])
|
||||
##plt.plot(region_sum_0)
|
||||
##plt.plot(peaks,region_sum_0[peaks],'*')
|
||||
##plt.show()
|
||||
#print(first_nonzero,last_nonzero,peaks)
|
||||
peaks=peaks[(peaks>first_nonzero) & ((peaks<last_nonzero))]
|
||||
|
||||
#print(first_nonzero,last_nonzero,peaks)
|
||||
|
||||
|
||||
#print(region_sum_0[peaks]<10)
|
||||
####peaks=peaks[region_sum_0[peaks]<25 ]
|
||||
|
||||
#print(region_sum_0[peaks])
|
||||
peaks=peaks[region_sum_0[peaks]<min_textline_thickness ]
|
||||
#print(peaks)
|
||||
#print(first_nonzero,last_nonzero,one_third_right,one_third_left)
|
||||
|
||||
if num_col==1:
|
||||
peaks_right=peaks[peaks>mid_point]
|
||||
peaks_left=peaks[peaks<mid_point]
|
||||
if num_col==2:
|
||||
peaks_right=peaks[peaks>(mid_point+one_third_right)]
|
||||
peaks_left=peaks[peaks<(mid_point-one_third_left)]
|
||||
|
||||
|
||||
try:
|
||||
point_right=np.min(peaks_right)
|
||||
except:
|
||||
point_right=last_nonzero
|
||||
|
||||
|
||||
try:
|
||||
point_left=np.max(peaks_left)
|
||||
except:
|
||||
point_left=first_nonzero
|
||||
|
||||
|
||||
|
||||
|
||||
#print(point_left,point_right)
|
||||
#print(text_regions.shape)
|
||||
if point_right>=mask_marginals.shape[1]:
|
||||
point_right=mask_marginals.shape[1]-1
|
||||
|
||||
try:
|
||||
mask_marginals[:,point_left:point_right]=1
|
||||
except:
|
||||
mask_marginals[:,:]=1
|
||||
|
||||
#print(mask_marginals.shape,point_left,point_right,'nadosh')
|
||||
mask_marginals_rotated=rotate_image(mask_marginals,-slope_deskew)
|
||||
|
||||
#print(mask_marginals_rotated.shape,'nadosh')
|
||||
mask_marginals_rotated_sum=mask_marginals_rotated.sum(axis=0)
|
||||
|
||||
mask_marginals_rotated_sum[mask_marginals_rotated_sum!=0]=1
|
||||
index_x=np.array(range(len(mask_marginals_rotated_sum)))+1
|
||||
|
||||
index_x_interest=index_x[mask_marginals_rotated_sum==1]
|
||||
|
||||
min_point_of_left_marginal=np.min(index_x_interest)-16
|
||||
max_point_of_right_marginal=np.max(index_x_interest)+16
|
||||
|
||||
if min_point_of_left_marginal<0:
|
||||
min_point_of_left_marginal=0
|
||||
if max_point_of_right_marginal>=text_regions.shape[1]:
|
||||
max_point_of_right_marginal=text_regions.shape[1]-1
|
||||
|
||||
|
||||
#print(np.min(index_x_interest) ,np.max(index_x_interest),'minmaxnew')
|
||||
#print(mask_marginals_rotated.shape,text_regions.shape,'mask_marginals_rotated')
|
||||
#plt.imshow(mask_marginals)
|
||||
#plt.show()
|
||||
|
||||
#plt.imshow(mask_marginals_rotated)
|
||||
#plt.show()
|
||||
|
||||
text_regions[(mask_marginals_rotated[:,:]!=1) & (text_regions[:,:]==1)]=4
|
||||
|
||||
#plt.imshow(text_regions)
|
||||
#plt.show()
|
||||
|
||||
pixel_img=4
|
||||
min_area_text=0.00001
|
||||
polygons_of_marginals=return_contours_of_interested_region(text_regions,pixel_img,min_area_text)
|
||||
|
||||
cx_text_only,cy_text_only ,x_min_text_only,x_max_text_only, y_min_text_only ,y_max_text_only,y_cor_x_min_main=find_new_features_of_contours(polygons_of_marginals)
|
||||
|
||||
text_regions[(text_regions[:,:]==4)]=1
|
||||
|
||||
marginlas_should_be_main_text=[]
|
||||
|
||||
x_min_marginals_left=[]
|
||||
x_min_marginals_right=[]
|
||||
|
||||
for i in range(len(cx_text_only)):
|
||||
|
||||
x_width_mar=abs(x_min_text_only[i]-x_max_text_only[i])
|
||||
y_height_mar=abs(y_min_text_only[i]-y_max_text_only[i])
|
||||
#print(x_width_mar,y_height_mar,y_height_mar/x_width_mar,'y_height_mar')
|
||||
if x_width_mar>16 and y_height_mar/x_width_mar<18:
|
||||
marginlas_should_be_main_text.append(polygons_of_marginals[i])
|
||||
if x_min_text_only[i]<(mid_point-one_third_left):
|
||||
x_min_marginals_left_new=x_min_text_only[i]
|
||||
if len(x_min_marginals_left)==0:
|
||||
x_min_marginals_left.append(x_min_marginals_left_new)
|
||||
else:
|
||||
x_min_marginals_left[0]=min(x_min_marginals_left[0],x_min_marginals_left_new)
|
||||
else:
|
||||
x_min_marginals_right_new=x_min_text_only[i]
|
||||
if len(x_min_marginals_right)==0:
|
||||
x_min_marginals_right.append(x_min_marginals_right_new)
|
||||
else:
|
||||
x_min_marginals_right[0]=min(x_min_marginals_right[0],x_min_marginals_right_new)
|
||||
|
||||
if len(x_min_marginals_left)==0:
|
||||
x_min_marginals_left=[0]
|
||||
if len(x_min_marginals_right)==0:
|
||||
x_min_marginals_right=[text_regions.shape[1]-1]
|
||||
|
||||
|
||||
|
||||
|
||||
#print(x_min_marginals_left[0],x_min_marginals_right[0],'margo')
|
||||
|
||||
#print(marginlas_should_be_main_text,'marginlas_should_be_main_text')
|
||||
text_regions=cv2.fillPoly(text_regions, pts =marginlas_should_be_main_text, color=(4,4))
|
||||
|
||||
#print(np.unique(text_regions))
|
||||
|
||||
#text_regions[:,:int(x_min_marginals_left[0])][text_regions[:,:int(x_min_marginals_left[0])]==1]=0
|
||||
#text_regions[:,int(x_min_marginals_right[0]):][text_regions[:,int(x_min_marginals_right[0]):]==1]=0
|
||||
|
||||
text_regions[:,:int(min_point_of_left_marginal)][text_regions[:,:int(min_point_of_left_marginal)]==1]=0
|
||||
text_regions[:,int(max_point_of_right_marginal):][text_regions[:,int(max_point_of_right_marginal):]==1]=0
|
||||
|
||||
###text_regions[:,0:point_left][text_regions[:,0:point_left]==1]=4
|
||||
|
||||
###text_regions[:,point_right:][ text_regions[:,point_right:]==1]=4
|
||||
#plt.plot(region_sum_0)
|
||||
#plt.plot(peaks,region_sum_0[peaks],'*')
|
||||
#plt.show()
|
||||
|
||||
|
||||
#plt.imshow(text_regions)
|
||||
#plt.show()
|
||||
|
||||
#sys.exit()
|
||||
else:
|
||||
pass
|
||||
return text_regions
|
||||
|
|
@ -1,4 +1,3 @@
|
|||
from contextlib import nullcontext
|
||||
from PIL import Image
|
||||
import numpy as np
|
||||
from ocrd_models import OcrdExif
|
||||
|
|
@ -17,13 +16,12 @@ def pil2cv(img):
|
|||
|
||||
def check_dpi(img):
|
||||
try:
|
||||
if isinstance(img, Image.Image):
|
||||
pil_image = nullcontext(img)
|
||||
if isinstance(img, Image.__class__):
|
||||
pil_image = img
|
||||
elif isinstance(img, str):
|
||||
pil_image = Image.open(img)
|
||||
else:
|
||||
pil_image = nullcontext(cv2pil(img))
|
||||
with pil_image:
|
||||
pil_image = cv2pil(img)
|
||||
exif = OcrdExif(pil_image)
|
||||
resolution = exif.resolution
|
||||
if resolution == 1:
|
||||
85
qurator/eynollah/utils/rotate.py
Normal file
85
qurator/eynollah/utils/rotate.py
Normal file
|
|
@ -0,0 +1,85 @@
|
|||
import math
|
||||
|
||||
import imutils
|
||||
import cv2
|
||||
|
||||
def rotatedRectWithMaxArea(w, h, angle):
|
||||
if w <= 0 or h <= 0:
|
||||
return 0, 0
|
||||
|
||||
width_is_longer = w >= h
|
||||
side_long, side_short = (w, h) if width_is_longer else (h, w)
|
||||
|
||||
# since the solutions for angle, -angle and 180-angle are all the same,
|
||||
# if suffices to look at the first quadrant and the absolute values of sin,cos:
|
||||
sin_a, cos_a = abs(math.sin(angle)), abs(math.cos(angle))
|
||||
if side_short <= 2.0 * sin_a * cos_a * side_long or abs(sin_a - cos_a) < 1e-10:
|
||||
# half constrained case: two crop corners touch the longer side,
|
||||
# the other two corners are on the mid-line parallel to the longer line
|
||||
x = 0.5 * side_short
|
||||
wr, hr = (x / sin_a, x / cos_a) if width_is_longer else (x / cos_a, x / sin_a)
|
||||
else:
|
||||
# fully constrained case: crop touches all 4 sides
|
||||
cos_2a = cos_a * cos_a - sin_a * sin_a
|
||||
wr, hr = (w * cos_a - h * sin_a) / cos_2a, (h * cos_a - w * sin_a) / cos_2a
|
||||
|
||||
return wr, hr
|
||||
|
||||
def rotate_max_area_new(image, rotated, angle):
|
||||
wr, hr = rotatedRectWithMaxArea(image.shape[1], image.shape[0], math.radians(angle))
|
||||
h, w, _ = rotated.shape
|
||||
y1 = h // 2 - int(hr / 2)
|
||||
y2 = y1 + int(hr)
|
||||
x1 = w // 2 - int(wr / 2)
|
||||
x2 = x1 + int(wr)
|
||||
return rotated[y1:y2, x1:x2]
|
||||
|
||||
def rotation_image_new(img, thetha):
|
||||
rotated = imutils.rotate(img, thetha)
|
||||
return rotate_max_area_new(img, rotated, thetha)
|
||||
|
||||
def rotate_image(img_patch, slope):
|
||||
(h, w) = img_patch.shape[:2]
|
||||
center = (w // 2, h // 2)
|
||||
M = cv2.getRotationMatrix2D(center, slope, 1.0)
|
||||
return cv2.warpAffine(img_patch, M, (w, h), flags=cv2.INTER_CUBIC, borderMode=cv2.BORDER_REPLICATE)
|
||||
|
||||
def rotate_image_different( img, slope):
|
||||
# img = cv2.imread('images/input.jpg')
|
||||
num_rows, num_cols = img.shape[:2]
|
||||
|
||||
rotation_matrix = cv2.getRotationMatrix2D((num_cols / 2, num_rows / 2), slope, 1)
|
||||
img_rotation = cv2.warpAffine(img, rotation_matrix, (num_cols, num_rows))
|
||||
return img_rotation
|
||||
|
||||
def rotate_max_area(image, rotated, rotated_textline, rotated_layout, angle):
|
||||
wr, hr = rotatedRectWithMaxArea(image.shape[1], image.shape[0], math.radians(angle))
|
||||
h, w, _ = rotated.shape
|
||||
y1 = h // 2 - int(hr / 2)
|
||||
y2 = y1 + int(hr)
|
||||
x1 = w // 2 - int(wr / 2)
|
||||
x2 = x1 + int(wr)
|
||||
return rotated[y1:y2, x1:x2], rotated_textline[y1:y2, x1:x2], rotated_layout[y1:y2, x1:x2]
|
||||
|
||||
def rotation_not_90_func(img, textline, text_regions_p_1, thetha):
|
||||
rotated = imutils.rotate(img, thetha)
|
||||
rotated_textline = imutils.rotate(textline, thetha)
|
||||
rotated_layout = imutils.rotate(text_regions_p_1, thetha)
|
||||
return rotate_max_area(img, rotated, rotated_textline, rotated_layout, thetha)
|
||||
|
||||
def rotation_not_90_func_full_layout(img, textline, text_regions_p_1, text_regions_p_fully, thetha):
|
||||
rotated = imutils.rotate(img, thetha)
|
||||
rotated_textline = imutils.rotate(textline, thetha)
|
||||
rotated_layout = imutils.rotate(text_regions_p_1, thetha)
|
||||
rotated_layout_full = imutils.rotate(text_regions_p_fully, thetha)
|
||||
return rotate_max_area_full_layout(img, rotated, rotated_textline, rotated_layout, rotated_layout_full, thetha)
|
||||
|
||||
def rotate_max_area_full_layout(image, rotated, rotated_textline, rotated_layout, rotated_layout_full, angle):
|
||||
wr, hr = rotatedRectWithMaxArea(image.shape[1], image.shape[0], math.radians(angle))
|
||||
h, w, _ = rotated.shape
|
||||
y1 = h // 2 - int(hr / 2)
|
||||
y2 = y1 + int(hr)
|
||||
x1 = w // 2 - int(wr / 2)
|
||||
x2 = x1 + int(wr)
|
||||
return rotated[y1:y2, x1:x2], rotated_textline[y1:y2, x1:x2], rotated_layout[y1:y2, x1:x2], rotated_layout_full[y1:y2, x1:x2]
|
||||
|
||||
File diff suppressed because it is too large
Load diff
|
|
@ -21,6 +21,7 @@ from ocrd_models.ocrd_page import (
|
|||
RegionRefType,
|
||||
SeparatorRegionType,
|
||||
TableRegionType,
|
||||
TextEquivType,
|
||||
TextLineType,
|
||||
TextRegionType,
|
||||
UnorderedGroupIndexedType,
|
||||
|
|
@ -46,26 +47,24 @@ def create_page_xml(imageFilename, height, width):
|
|||
))
|
||||
return pcgts
|
||||
|
||||
def xml_reading_order(page, order_of_texts, id_of_marginalia_left, id_of_marginalia_right):
|
||||
def xml_reading_order(page, order_of_texts, id_of_marginalia):
|
||||
region_order = ReadingOrderType()
|
||||
og = OrderedGroupType(id="ro357564684568544579089")
|
||||
page.set_ReadingOrder(region_order)
|
||||
region_order.set_OrderedGroup(og)
|
||||
region_counter = EynollahIdCounter()
|
||||
|
||||
for id_marginal in id_of_marginalia_left:
|
||||
for idx_textregion, _ in enumerate(order_of_texts):
|
||||
og.add_RegionRefIndexed(RegionRefIndexedType(index=str(region_counter.get('region')), regionRef=region_counter.region_id(order_of_texts[idx_textregion] + 1)))
|
||||
region_counter.inc('region')
|
||||
for id_marginal in id_of_marginalia:
|
||||
og.add_RegionRefIndexed(RegionRefIndexedType(index=str(region_counter.get('region')), regionRef=id_marginal))
|
||||
region_counter.inc('region')
|
||||
|
||||
for idx_textregion in order_of_texts:
|
||||
og.add_RegionRefIndexed(RegionRefIndexedType(index=str(region_counter.get('region')), regionRef=region_counter.region_id(idx_textregion + 1)))
|
||||
region_counter.inc('region')
|
||||
def order_and_id_of_texts(found_polygons_text_region, found_polygons_text_region_h, matrix_of_orders, indexes_sorted, index_of_types, kind_of_texts, ref_point):
|
||||
indexes_sorted = np.array(indexes_sorted)
|
||||
index_of_types = np.array(index_of_types)
|
||||
kind_of_texts = np.array(kind_of_texts)
|
||||
|
||||
for id_marginal in id_of_marginalia_right:
|
||||
og.add_RegionRefIndexed(RegionRefIndexedType(index=str(region_counter.get('region')), regionRef=id_marginal))
|
||||
region_counter.inc('region')
|
||||
|
||||
def order_and_id_of_texts(found_polygons_text_region, found_polygons_text_region_h, indexes_sorted, index_of_types, kind_of_texts, ref_point):
|
||||
id_of_texts = []
|
||||
order_of_texts = []
|
||||
|
||||
|
|
@ -88,7 +87,3 @@ def order_and_id_of_texts(found_polygons_text_region, found_polygons_text_region
|
|||
order_of_texts.append(interest)
|
||||
|
||||
return order_of_texts, id_of_texts
|
||||
|
||||
def etree_namespace_for_element_tag(tag: str):
|
||||
right = tag.find('}')
|
||||
return tag[1:right]
|
||||
260
qurator/eynollah/writer.py
Normal file
260
qurator/eynollah/writer.py
Normal file
|
|
@ -0,0 +1,260 @@
|
|||
# pylint: disable=too-many-locals,wrong-import-position,too-many-lines,too-many-statements,chained-comparison,fixme,broad-except,c-extension-no-member
|
||||
# pylint: disable=import-error
|
||||
from pathlib import Path
|
||||
import os.path
|
||||
|
||||
from .utils.xml import create_page_xml, xml_reading_order
|
||||
from .utils.counter import EynollahIdCounter
|
||||
|
||||
from ocrd_utils import getLogger
|
||||
from ocrd_models.ocrd_page import (
|
||||
BorderType,
|
||||
CoordsType,
|
||||
TextEquivType,
|
||||
PcGtsType,
|
||||
TextLineType,
|
||||
TextRegionType,
|
||||
ImageRegionType,
|
||||
TableRegionType,
|
||||
SeparatorRegionType,
|
||||
to_xml
|
||||
)
|
||||
import numpy as np
|
||||
|
||||
class EynollahXmlWriter():
|
||||
|
||||
def __init__(self, *, dir_out, image_filename, curved_line, pcgts=None):
|
||||
self.logger = getLogger('eynollah.writer')
|
||||
self.counter = EynollahIdCounter()
|
||||
self.dir_out = dir_out
|
||||
self.image_filename = image_filename
|
||||
self.curved_line = curved_line
|
||||
self.pcgts = pcgts
|
||||
self.scale_x = None # XXX set outside __init__
|
||||
self.scale_y = None # XXX set outside __init__
|
||||
self.height_org = None # XXX set outside __init__
|
||||
self.width_org = None # XXX set outside __init__
|
||||
|
||||
@property
|
||||
def image_filename_stem(self):
|
||||
return Path(Path(self.image_filename).name).stem
|
||||
|
||||
def calculate_page_coords(self, cont_page):
|
||||
self.logger.debug('enter calculate_page_coords')
|
||||
points_page_print = ""
|
||||
for _, contour in enumerate(cont_page[0]):
|
||||
if len(contour) == 2:
|
||||
points_page_print += str(int((contour[0]) / self.scale_x))
|
||||
points_page_print += ','
|
||||
points_page_print += str(int((contour[1]) / self.scale_y))
|
||||
else:
|
||||
points_page_print += str(int((contour[0][0]) / self.scale_x))
|
||||
points_page_print += ','
|
||||
points_page_print += str(int((contour[0][1] ) / self.scale_y))
|
||||
points_page_print = points_page_print + ' '
|
||||
return points_page_print[:-1]
|
||||
|
||||
def serialize_lines_in_marginal(self, marginal_region, all_found_texline_polygons_marginals, marginal_idx, page_coord, all_box_coord_marginals, slopes_marginals, counter):
|
||||
for j in range(len(all_found_texline_polygons_marginals[marginal_idx])):
|
||||
coords = CoordsType()
|
||||
textline = TextLineType(id=counter.next_line_id, Coords=coords)
|
||||
marginal_region.add_TextLine(textline)
|
||||
textline.add_TextEquiv(TextEquivType(Unicode=''))
|
||||
points_co = ''
|
||||
for l in range(len(all_found_texline_polygons_marginals[marginal_idx][j])):
|
||||
if not self.curved_line:
|
||||
if len(all_found_texline_polygons_marginals[marginal_idx][j][l]) == 2:
|
||||
textline_x_coord = max(0, int((all_found_texline_polygons_marginals[marginal_idx][j][l][0] + all_box_coord_marginals[marginal_idx][2] + page_coord[2]) / self.scale_x) )
|
||||
textline_y_coord = max(0, int((all_found_texline_polygons_marginals[marginal_idx][j][l][1] + all_box_coord_marginals[marginal_idx][0] + page_coord[0]) / self.scale_y) )
|
||||
else:
|
||||
textline_x_coord = max(0, int((all_found_texline_polygons_marginals[marginal_idx][j][l][0][0] + all_box_coord_marginals[marginal_idx][2] + page_coord[2]) / self.scale_x) )
|
||||
textline_y_coord = max(0, int((all_found_texline_polygons_marginals[marginal_idx][j][l][0][1] + all_box_coord_marginals[marginal_idx][0] + page_coord[0]) / self.scale_y) )
|
||||
points_co += str(textline_x_coord)
|
||||
points_co += ','
|
||||
points_co += str(textline_y_coord)
|
||||
if self.curved_line and np.abs(slopes_marginals[marginal_idx]) <= 45:
|
||||
if len(all_found_texline_polygons_marginals[marginal_idx][j][l]) == 2:
|
||||
points_co += str(int((all_found_texline_polygons_marginals[marginal_idx][j][l][0] + page_coord[2]) / self.scale_x))
|
||||
points_co += ','
|
||||
points_co += str(int((all_found_texline_polygons_marginals[marginal_idx][j][l][1] + page_coord[0]) / self.scale_y))
|
||||
else:
|
||||
points_co += str(int((all_found_texline_polygons_marginals[marginal_idx][j][l][0][0] + page_coord[2]) / self.scale_x))
|
||||
points_co += ','
|
||||
points_co += str(int((all_found_texline_polygons_marginals[marginal_idx][j][l][0][1] + page_coord[0]) / self.scale_y))
|
||||
|
||||
elif self.curved_line and np.abs(slopes_marginals[marginal_idx]) > 45:
|
||||
if len(all_found_texline_polygons_marginals[marginal_idx][j][l]) == 2:
|
||||
points_co += str(int((all_found_texline_polygons_marginals[marginal_idx][j][l][0] + all_box_coord_marginals[marginal_idx][2] + page_coord[2]) / self.scale_x))
|
||||
points_co += ','
|
||||
points_co += str(int((all_found_texline_polygons_marginals[marginal_idx][j][l][1] + all_box_coord_marginals[marginal_idx][0] + page_coord[0]) / self.scale_y))
|
||||
else:
|
||||
points_co += str(int((all_found_texline_polygons_marginals[marginal_idx][j][l][0][0] + all_box_coord_marginals[marginal_idx][2] + page_coord[2]) / self.scale_x))
|
||||
points_co += ','
|
||||
points_co += str(int((all_found_texline_polygons_marginals[marginal_idx][j][l][0][1] + all_box_coord_marginals[marginal_idx][0] + page_coord[0]) / self.scale_y))
|
||||
points_co += ' '
|
||||
coords.set_points(points_co[:-1])
|
||||
|
||||
def serialize_lines_in_region(self, text_region, all_found_texline_polygons, region_idx, page_coord, all_box_coord, slopes, counter):
|
||||
self.logger.debug('enter serialize_lines_in_region')
|
||||
for j in range(len(all_found_texline_polygons[region_idx])):
|
||||
coords = CoordsType()
|
||||
textline = TextLineType(id=counter.next_line_id, Coords=coords, TextEquiv=[TextEquivType(index=0, Unicode='')])
|
||||
text_region.add_TextLine(textline)
|
||||
region_bboxes = all_box_coord[region_idx]
|
||||
points_co = ''
|
||||
for idx_contour_textline, contour_textline in enumerate(all_found_texline_polygons[region_idx][j]):
|
||||
if not self.curved_line:
|
||||
if len(contour_textline) == 2:
|
||||
textline_x_coord = max(0, int((contour_textline[0] + region_bboxes[2] + page_coord[2]) / self.scale_x))
|
||||
textline_y_coord = max(0, int((contour_textline[1] + region_bboxes[0] + page_coord[0]) / self.scale_y))
|
||||
else:
|
||||
textline_x_coord = max(0, int((contour_textline[0][0] + region_bboxes[2] + page_coord[2]) / self.scale_x))
|
||||
textline_y_coord = max(0, int((contour_textline[0][1] + region_bboxes[0] + page_coord[0]) / self.scale_y))
|
||||
points_co += str(textline_x_coord)
|
||||
points_co += ','
|
||||
points_co += str(textline_y_coord)
|
||||
|
||||
if self.curved_line and np.abs(slopes[region_idx]) <= 45:
|
||||
if len(contour_textline) == 2:
|
||||
points_co += str(int((contour_textline[0] + page_coord[2]) / self.scale_x))
|
||||
points_co += ','
|
||||
points_co += str(int((contour_textline[1] + page_coord[0]) / self.scale_y))
|
||||
else:
|
||||
points_co += str(int((contour_textline[0][0] + page_coord[2]) / self.scale_x))
|
||||
points_co += ','
|
||||
points_co += str(int((contour_textline[0][1] + page_coord[0])/self.scale_y))
|
||||
elif self.curved_line and np.abs(slopes[region_idx]) > 45:
|
||||
if len(contour_textline)==2:
|
||||
points_co += str(int((contour_textline[0] + region_bboxes[2] + page_coord[2])/self.scale_x))
|
||||
points_co += ','
|
||||
points_co += str(int((contour_textline[1] + region_bboxes[0] + page_coord[0])/self.scale_y))
|
||||
else:
|
||||
points_co += str(int((contour_textline[0][0] + region_bboxes[2]+page_coord[2])/self.scale_x))
|
||||
points_co += ','
|
||||
points_co += str(int((contour_textline[0][1] + region_bboxes[0]+page_coord[0])/self.scale_y))
|
||||
points_co += ' '
|
||||
coords.set_points(points_co[:-1])
|
||||
|
||||
def write_pagexml(self, pcgts):
|
||||
out_fname = os.path.join(self.dir_out, self.image_filename_stem) + ".xml"
|
||||
self.logger.info("output filename: '%s'", out_fname)
|
||||
with open(out_fname, 'w') as f:
|
||||
f.write(to_xml(pcgts))
|
||||
|
||||
def build_pagexml_no_full_layout(self, found_polygons_text_region, page_coord, order_of_texts, id_of_texts, all_found_texline_polygons, all_box_coord, found_polygons_text_region_img, found_polygons_marginals, all_found_texline_polygons_marginals, all_box_coord_marginals, slopes, slopes_marginals, cont_page, polygons_lines_to_be_written_in_xml):
|
||||
self.logger.debug('enter build_pagexml_no_full_layout')
|
||||
|
||||
# create the file structure
|
||||
pcgts = self.pcgts if self.pcgts else create_page_xml(self.image_filename, self.height_org, self.width_org)
|
||||
page = pcgts.get_Page()
|
||||
page.set_Border(BorderType(Coords=CoordsType(points=self.calculate_page_coords(cont_page))))
|
||||
|
||||
counter = EynollahIdCounter()
|
||||
if len(found_polygons_text_region) > 0:
|
||||
_counter_marginals = EynollahIdCounter(region_idx=len(order_of_texts))
|
||||
id_of_marginalia = [_counter_marginals.next_region_id for _ in found_polygons_marginals]
|
||||
xml_reading_order(page, order_of_texts, id_of_marginalia)
|
||||
|
||||
for mm in range(len(found_polygons_text_region)):
|
||||
textregion = TextRegionType(id=counter.next_region_id, type_='paragraph',
|
||||
Coords=CoordsType(points=self.calculate_polygon_coords(found_polygons_text_region[mm], page_coord)),
|
||||
TextEquiv=[TextEquivType(index=0, Unicode='')])
|
||||
page.add_TextRegion(textregion)
|
||||
self.serialize_lines_in_region(textregion, all_found_texline_polygons, mm, page_coord, all_box_coord, slopes, counter)
|
||||
|
||||
for mm in range(len(found_polygons_marginals)):
|
||||
marginal = TextRegionType(id=counter.next_region_id, type_='marginalia',
|
||||
Coords=CoordsType(points=self.calculate_polygon_coords(found_polygons_marginals[mm], page_coord)))
|
||||
page.add_TextRegion(marginal)
|
||||
self.serialize_lines_in_marginal(marginal, all_found_texline_polygons_marginals, mm, page_coord, all_box_coord_marginals, slopes_marginals, counter)
|
||||
|
||||
for mm in range(len(found_polygons_text_region_img)):
|
||||
img_region = ImageRegionType(id=counter.next_region_id, Coords=CoordsType())
|
||||
page.add_ImageRegion(img_region)
|
||||
points_co = ''
|
||||
for lmm in range(len(found_polygons_text_region_img[mm])):
|
||||
points_co += str(int((found_polygons_text_region_img[mm][lmm,0,0] + page_coord[2]) / self.scale_x))
|
||||
points_co += ','
|
||||
points_co += str(int((found_polygons_text_region_img[mm][lmm,0,1] + page_coord[0]) / self.scale_y))
|
||||
points_co += ' '
|
||||
img_region.get_Coords().set_points(points_co[:-1])
|
||||
|
||||
for mm in range(len(polygons_lines_to_be_written_in_xml)):
|
||||
sep_hor = SeparatorRegionType(id=counter.next_region_id, Coords=CoordsType())
|
||||
page.add_SeparatorRegion(sep_hor)
|
||||
points_co = ''
|
||||
for lmm in range(len(polygons_lines_to_be_written_in_xml[mm])):
|
||||
points_co += str(int((polygons_lines_to_be_written_in_xml[mm][lmm,0,0] ) / self.scale_x))
|
||||
points_co += ','
|
||||
points_co += str(int((polygons_lines_to_be_written_in_xml[mm][lmm,0,1] ) / self.scale_y))
|
||||
points_co += ' '
|
||||
sep_hor.get_Coords().set_points(points_co[:-1])
|
||||
|
||||
return pcgts
|
||||
|
||||
def build_pagexml_full_layout(self, found_polygons_text_region, found_polygons_text_region_h, page_coord, order_of_texts, id_of_texts, all_found_texline_polygons, all_found_texline_polygons_h, all_box_coord, all_box_coord_h, found_polygons_text_region_img, found_polygons_tables, found_polygons_drop_capitals, found_polygons_marginals, all_found_texline_polygons_marginals, all_box_coord_marginals, slopes, slopes_h, slopes_marginals, cont_page, polygons_lines_to_be_written_in_xml):
|
||||
self.logger.debug('enter build_pagexml_full_layout')
|
||||
|
||||
# create the file structure
|
||||
pcgts = self.pcgts if self.pcgts else create_page_xml(self.image_filename, self.height_org, self.width_org)
|
||||
page = pcgts.get_Page()
|
||||
page.set_Border(BorderType(Coords=CoordsType(points=self.calculate_page_coords(cont_page))))
|
||||
|
||||
counter = EynollahIdCounter()
|
||||
_counter_marginals = EynollahIdCounter(region_idx=len(order_of_texts))
|
||||
id_of_marginalia = [_counter_marginals.next_region_id for _ in found_polygons_marginals]
|
||||
xml_reading_order(page, order_of_texts, id_of_marginalia)
|
||||
|
||||
for mm in range(len(found_polygons_text_region)):
|
||||
textregion = TextRegionType(id=counter.next_region_id, type_='paragraph',
|
||||
TextEquiv=[TextEquivType(index=0, Unicode='')],
|
||||
Coords=CoordsType(points=self.calculate_polygon_coords(found_polygons_text_region[mm], page_coord)))
|
||||
page.add_TextRegion(textregion)
|
||||
self.serialize_lines_in_region(textregion, all_found_texline_polygons, mm, page_coord, all_box_coord, slopes, counter)
|
||||
|
||||
self.logger.debug('len(found_polygons_text_region_h) %s', len(found_polygons_text_region_h))
|
||||
for mm in range(len(found_polygons_text_region_h)):
|
||||
textregion = TextRegionType(id=counter.next_region_id, type_='header',
|
||||
TextEquiv=[TextEquivType(index=0, Unicode='')],
|
||||
Coords=CoordsType(points=self.calculate_polygon_coords(found_polygons_text_region_h[mm], page_coord)))
|
||||
page.add_TextRegion(textregion)
|
||||
self.serialize_lines_in_region(textregion, all_found_texline_polygons_h, mm, page_coord, all_box_coord_h, slopes_h, counter)
|
||||
|
||||
for mm in range(len(found_polygons_marginals)):
|
||||
marginal = TextRegionType(id=counter.next_region_id, type_='marginalia',
|
||||
TextEquiv=[TextEquivType(index=0, Unicode='')],
|
||||
Coords=CoordsType(points=self.calculate_polygon_coords(found_polygons_marginals[mm], page_coord)))
|
||||
page.add_TextRegion(marginal)
|
||||
self.serialize_lines_in_marginal(marginal, all_found_texline_polygons_marginals, mm, page_coord, all_box_coord_marginals, slopes_marginals, counter)
|
||||
|
||||
for mm in range(len(found_polygons_drop_capitals)):
|
||||
page.add_TextRegion(TextRegionType(id=counter.next_region_id, type_='drop-capital',
|
||||
TextEquiv=[TextEquivType(index=0, Unicode='')],
|
||||
Coords=CoordsType(points=self.calculate_polygon_coords(found_polygons_drop_capitals[mm], page_coord))))
|
||||
|
||||
for mm in range(len(found_polygons_text_region_img)):
|
||||
page.add_ImageRegion(ImageRegionType(id=counter.next_region_id, Coords=CoordsType(points=self.calculate_polygon_coords(found_polygons_text_region_img[mm], page_coord))))
|
||||
|
||||
for mm in range(len(polygons_lines_to_be_written_in_xml)):
|
||||
page.add_SeparatorRegion(ImageRegionType(id=counter.next_region_id, Coords=CoordsType(points=self.calculate_polygon_coords(polygons_lines_to_be_written_in_xml[mm], [0 , 0, 0, 0]))))
|
||||
|
||||
for mm in range(len(found_polygons_tables)):
|
||||
page.add_TableRegion(TableRegionType(id=counter.next_region_id, Coords=CoordsType(points=self.calculate_polygon_coords(found_polygons_tables[mm], page_coord))))
|
||||
|
||||
return pcgts
|
||||
|
||||
def calculate_polygon_coords(self, contour, page_coord):
|
||||
self.logger.debug('enter calculate_polygon_coords')
|
||||
coords = ''
|
||||
for value_bbox in contour:
|
||||
if len(value_bbox) == 2:
|
||||
coords += str(int((value_bbox[0] + page_coord[2]) / self.scale_x))
|
||||
coords += ','
|
||||
coords += str(int((value_bbox[1] + page_coord[0]) / self.scale_y))
|
||||
else:
|
||||
coords += str(int((value_bbox[0][0] + page_coord[2]) / self.scale_x))
|
||||
coords += ','
|
||||
coords += str(int((value_bbox[0][1] + page_coord[0]) / self.scale_y))
|
||||
coords=coords + ' '
|
||||
return coords[:-1]
|
||||
|
||||
|
|
@ -1,3 +0,0 @@
|
|||
torch
|
||||
transformers <= 4.30.2 ; python_version < '3.10'
|
||||
transformers >= 5 ; python_version >= '3.10'
|
||||
|
|
@ -1 +0,0 @@
|
|||
matplotlib
|
||||
|
|
@ -1,4 +1,2 @@
|
|||
pytest
|
||||
pytest-isolate
|
||||
coverage[toml]
|
||||
black
|
||||
|
|
|
|||
|
|
@ -1 +0,0 @@
|
|||
train/requirements.txt
|
||||
|
|
@ -1,9 +1,8 @@
|
|||
# ocrd includes opencv, numpy, shapely, click
|
||||
ocrd >= 3.3.0
|
||||
numpy < 2.0
|
||||
ocrd >= 2.23.3
|
||||
keras >= 2.3.1, < 2.4
|
||||
scikit-learn >= 0.23.2
|
||||
tensorflow
|
||||
tf-keras # avoid keras 3 (also needs TF_USE_LEGACY_KERAS=1)
|
||||
numba <= 0.58.1
|
||||
scikit-image
|
||||
tabulate
|
||||
tensorflow-gpu >= 1.15, < 2
|
||||
imutils >= 0.5.3
|
||||
matplotlib
|
||||
setuptools >= 50
|
||||
|
|
|
|||
28
setup.py
Normal file
28
setup.py
Normal file
|
|
@ -0,0 +1,28 @@
|
|||
from setuptools import setup, find_packages
|
||||
from json import load
|
||||
|
||||
install_requires = open('requirements.txt').read().split('\n')
|
||||
with open('ocrd-tool.json', 'r', encoding='utf-8') as f:
|
||||
version = load(f)['version']
|
||||
|
||||
setup(
|
||||
name='eynollah',
|
||||
version=version,
|
||||
long_description=open('README.md').read(),
|
||||
long_description_content_type='text/markdown',
|
||||
author='Vahid Rezanezhad',
|
||||
url='https://github.com/qurator-spk/eynollah',
|
||||
license='Apache License 2.0',
|
||||
namespace_packages=['qurator'],
|
||||
packages=find_packages(exclude=['tests']),
|
||||
install_requires=install_requires,
|
||||
package_data={
|
||||
'': ['*.json']
|
||||
},
|
||||
entry_points={
|
||||
'console_scripts': [
|
||||
'eynollah=qurator.eynollah.cli:main',
|
||||
'ocrd-eynollah-segment=qurator.eynollah.ocrd_cli:main',
|
||||
]
|
||||
},
|
||||
)
|
||||
Binary file not shown.
|
|
@ -1,20 +0,0 @@
|
|||
# NOTE: For predictable order of imports of torch/shapely/tensorflow
|
||||
# this must be the first import of the CLI!
|
||||
from ..eynollah_imports import imported_libs
|
||||
|
||||
from .cli import main
|
||||
from .cli_binarize import binarize_cli
|
||||
from .cli_enhance import enhance_cli
|
||||
from .cli_extract_images import extract_images_cli
|
||||
from .cli_layout import layout_cli
|
||||
from .cli_models import models_cli
|
||||
from .cli_ocr import ocr_cli
|
||||
from .cli_readingorder import readingorder_cli
|
||||
|
||||
main.add_command(binarize_cli, 'binarization')
|
||||
main.add_command(enhance_cli, 'enhancement')
|
||||
main.add_command(layout_cli, 'layout')
|
||||
main.add_command(readingorder_cli, 'machine-based-reading-order')
|
||||
main.add_command(models_cli, 'models')
|
||||
main.add_command(ocr_cli, 'ocr')
|
||||
main.add_command(extract_images_cli, 'extract-images')
|
||||
|
|
@ -1,66 +0,0 @@
|
|||
from dataclasses import dataclass
|
||||
import logging
|
||||
import sys
|
||||
import os
|
||||
from typing import Union
|
||||
|
||||
import click
|
||||
|
||||
from ..model_zoo import EynollahModelZoo
|
||||
from .cli_models import models_cli
|
||||
|
||||
@dataclass()
|
||||
class EynollahCliCtx:
|
||||
"""
|
||||
Holds options relevant for all eynollah subcommands
|
||||
"""
|
||||
model_zoo: EynollahModelZoo
|
||||
log_level : Union[str, None] = 'INFO'
|
||||
|
||||
|
||||
@click.group()
|
||||
@click.option(
|
||||
"--model-basedir",
|
||||
"-m",
|
||||
help="directory of models",
|
||||
# NOTE: not mandatory to exist so --help for subcommands works but will log a warning
|
||||
# and raise exception when trying to load models in the CLI
|
||||
# type=click.Path(exists=True),
|
||||
default=os.getcwd(),
|
||||
)
|
||||
@click.option(
|
||||
"--model-overrides",
|
||||
"-mv",
|
||||
help="override default versions of model categories, syntax is 'CATEGORY VARIANT PATH', e.g 'region light /path/to/model'. See eynollah list-models for the full list",
|
||||
type=(str, str, str),
|
||||
multiple=True,
|
||||
)
|
||||
@click.option(
|
||||
"--log_level",
|
||||
"-l",
|
||||
type=click.Choice(['OFF', 'DEBUG', 'INFO', 'WARN', 'ERROR']),
|
||||
help="Override log level globally to this",
|
||||
)
|
||||
@click.pass_context
|
||||
def main(ctx, model_basedir, model_overrides, log_level):
|
||||
"""
|
||||
eynollah - Document Layout Analysis, Image Enhancement, OCR
|
||||
"""
|
||||
# Initialize logging
|
||||
console_handler = logging.StreamHandler(sys.stderr)
|
||||
console_handler.setLevel(logging.NOTSET)
|
||||
formatter = logging.Formatter('%(asctime)s.%(msecs)03d %(levelname)s %(name)s - %(message)s', datefmt='%H:%M:%S')
|
||||
console_handler.setFormatter(formatter)
|
||||
logging.getLogger('eynollah').addHandler(console_handler)
|
||||
logging.getLogger('eynollah').setLevel(log_level or logging.INFO)
|
||||
# Initialize model zoo
|
||||
model_zoo = EynollahModelZoo(basedir=model_basedir, model_overrides=model_overrides)
|
||||
# Initialize CLI context
|
||||
ctx.obj = EynollahCliCtx(
|
||||
model_zoo=model_zoo,
|
||||
log_level=log_level,
|
||||
)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
|
|
@ -1,62 +0,0 @@
|
|||
import click
|
||||
|
||||
@click.command()
|
||||
@click.option(
|
||||
'--patches/--no-patches',
|
||||
default=True,
|
||||
help='let the model see the image in patches (tiling) instead of total (full).'
|
||||
)
|
||||
@click.option(
|
||||
"--input-image", "--image",
|
||||
"-i",
|
||||
help="input image filename",
|
||||
type=click.Path(exists=True, dir_okay=False)
|
||||
)
|
||||
@click.option(
|
||||
"--dir_in",
|
||||
"-di",
|
||||
help="directory of input images (instead of --image)",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
)
|
||||
@click.option(
|
||||
"--output",
|
||||
"-o",
|
||||
help="output image (if using -i) or output image directory (if using -di)",
|
||||
type=click.Path(file_okay=True, dir_okay=True),
|
||||
required=True,
|
||||
)
|
||||
@click.option(
|
||||
"--overwrite",
|
||||
"-O",
|
||||
help="overwrite (instead of skipping) if output xml exists",
|
||||
is_flag=True,
|
||||
)
|
||||
@click.option(
|
||||
"--device",
|
||||
"-D",
|
||||
help="placement of computations in predictors for each model type; if none (by default), will try to use first available GPU or fall back to CPU; set string to force using a device (e.g. 'GPU0', 'GPU1' or 'CPU'). Can also be a comma-separated list of model category to device mappings (e.g. 'col_classifier:CPU,page:GPU0,*:GPU1')",
|
||||
)
|
||||
@click.pass_context
|
||||
def binarize_cli(
|
||||
ctx,
|
||||
patches,
|
||||
input_image,
|
||||
dir_in,
|
||||
output,
|
||||
overwrite,
|
||||
device,
|
||||
):
|
||||
"""
|
||||
Binarize images with a ML model
|
||||
"""
|
||||
from ..sbb_binarize import SbbBinarizer
|
||||
assert bool(input_image) != bool(dir_in), "Either -i (single input) or -di (directory) must be provided, but not both."
|
||||
binarizer = SbbBinarizer(model_zoo=ctx.obj.model_zoo, device=device)
|
||||
binarizer.run(
|
||||
image_filename=input_image,
|
||||
use_patches=patches,
|
||||
output=output,
|
||||
dir_in=dir_in,
|
||||
overwrite=overwrite
|
||||
)
|
||||
|
||||
|
|
@ -1,73 +0,0 @@
|
|||
import click
|
||||
|
||||
@click.command()
|
||||
@click.option(
|
||||
"--image",
|
||||
"-i",
|
||||
help="input image filename",
|
||||
type=click.Path(exists=True, dir_okay=False),
|
||||
)
|
||||
@click.option(
|
||||
"--out",
|
||||
"-o",
|
||||
help="directory for output image files",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
required=True,
|
||||
)
|
||||
@click.option(
|
||||
"--overwrite",
|
||||
"-O",
|
||||
help="overwrite (instead of skipping) if output image exists",
|
||||
is_flag=True,
|
||||
)
|
||||
@click.option(
|
||||
"--dir_in",
|
||||
"-di",
|
||||
help="directory of input images (instead of --image)",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
)
|
||||
@click.option(
|
||||
"--num_col_upper",
|
||||
"-ncu",
|
||||
default=0,
|
||||
type=click.IntRange(min=0),
|
||||
help="lower limit of columns in document image",
|
||||
)
|
||||
@click.option(
|
||||
"--num_col_lower",
|
||||
"-ncl",
|
||||
default=0,
|
||||
type=click.IntRange(min=0),
|
||||
help="upper limit of columns in document image",
|
||||
)
|
||||
@click.option(
|
||||
"--save_org_scale",
|
||||
"-sos",
|
||||
is_flag=True,
|
||||
help="save the enhanced image in original image size",
|
||||
)
|
||||
@click.option(
|
||||
"--device",
|
||||
"-D",
|
||||
help="placement of computations in predictors for each model type; if none (by default), will try to use first available GPU or fall back to CPU; set string to force using a device (e.g. 'GPU0', 'GPU1' or 'CPU'). Can also be a comma-separated list of model category to device mappings (e.g. 'col_classifier:CPU,page:GPU0,*:GPU1')",
|
||||
)
|
||||
@click.pass_context
|
||||
def enhance_cli(ctx, image, out, overwrite, dir_in, num_col_upper, num_col_lower, save_org_scale, device):
|
||||
"""
|
||||
Enhance image
|
||||
"""
|
||||
assert bool(image) != bool(dir_in), "Either -i (single input) or -di (directory) must be provided, but not both."
|
||||
from ..image_enhancer import Enhancer
|
||||
enhancer = Enhancer(
|
||||
model_zoo=ctx.obj.model_zoo,
|
||||
num_col_upper=num_col_upper,
|
||||
num_col_lower=num_col_lower,
|
||||
save_org_scale=save_org_scale,
|
||||
device=device,
|
||||
)
|
||||
enhancer.run(overwrite=overwrite,
|
||||
dir_in=dir_in,
|
||||
image_filename=image,
|
||||
dir_out=out,
|
||||
)
|
||||
|
||||
|
|
@ -1,100 +0,0 @@
|
|||
import click
|
||||
|
||||
@click.command()
|
||||
@click.option(
|
||||
"--image",
|
||||
"-i",
|
||||
help="input image filename",
|
||||
type=click.Path(exists=True, dir_okay=False),
|
||||
)
|
||||
|
||||
@click.option(
|
||||
"--out",
|
||||
"-o",
|
||||
help="directory for output PAGE-XML files",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
required=True,
|
||||
)
|
||||
@click.option(
|
||||
"--overwrite",
|
||||
"-O",
|
||||
help="overwrite (instead of skipping) if output xml exists",
|
||||
is_flag=True,
|
||||
)
|
||||
@click.option(
|
||||
"--dir_in",
|
||||
"-di",
|
||||
help="directory of input images (instead of --image)",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
)
|
||||
@click.option(
|
||||
"--save_images",
|
||||
"-si",
|
||||
help="if a directory is given, images in documents will be cropped and saved there",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
)
|
||||
@click.option(
|
||||
"--enable-plotting/--disable-plotting",
|
||||
"-ep/-noep",
|
||||
is_flag=True,
|
||||
help="If set, will plot intermediary files and images",
|
||||
)
|
||||
@click.option(
|
||||
"--input_binary/--input-RGB",
|
||||
"-ib/-irgb",
|
||||
is_flag=True,
|
||||
help="In general, eynollah uses RGB as input but if the input document is very dark, very bright or for any other reason you can turn on input binarization. When this flag is set, eynollah will binarize the RGB input document, you should always provide RGB images to eynollah.",
|
||||
)
|
||||
@click.option(
|
||||
"--ignore_page_extraction/--extract_page_included",
|
||||
"-ipe/-epi",
|
||||
is_flag=True,
|
||||
help="if this parameter set to true, this tool would ignore page extraction",
|
||||
)
|
||||
@click.option(
|
||||
"--num_col_upper",
|
||||
"-ncu",
|
||||
help="lower limit of columns in document image",
|
||||
)
|
||||
@click.option(
|
||||
"--num_col_lower",
|
||||
"-ncl",
|
||||
help="upper limit of columns in document image",
|
||||
)
|
||||
@click.pass_context
|
||||
def extract_images_cli(
|
||||
ctx,
|
||||
image,
|
||||
out,
|
||||
overwrite,
|
||||
dir_in,
|
||||
save_images,
|
||||
enable_plotting,
|
||||
input_binary,
|
||||
num_col_upper,
|
||||
num_col_lower,
|
||||
ignore_page_extraction,
|
||||
):
|
||||
"""
|
||||
Detect Layout (with optional image enhancement and reading order detection)
|
||||
"""
|
||||
assert enable_plotting or not save_images, "Plotting with -si also requires -ep"
|
||||
assert not enable_plotting or save_images, "Plotting with -ep also requires -si"
|
||||
assert bool(image) != bool(dir_in), "Either -i (single input) or -di (directory) must be provided, but not both."
|
||||
|
||||
from ..extract_images import EynollahImageExtractor
|
||||
extractor = EynollahImageExtractor(
|
||||
model_zoo=ctx.obj.model_zoo,
|
||||
enable_plotting=enable_plotting,
|
||||
input_binary=input_binary,
|
||||
ignore_page_extraction=ignore_page_extraction,
|
||||
num_col_upper=num_col_upper,
|
||||
num_col_lower=num_col_lower,
|
||||
)
|
||||
extractor.run(overwrite=overwrite,
|
||||
image_filename=image,
|
||||
dir_in=dir_in,
|
||||
dir_out=out,
|
||||
dir_of_cropped_images=save_images,
|
||||
)
|
||||
|
||||
|
|
@ -1,256 +0,0 @@
|
|||
import click
|
||||
|
||||
@click.command(context_settings=dict(
|
||||
help_option_names=['-h', '--help'],
|
||||
show_default=True))
|
||||
@click.option(
|
||||
"--image",
|
||||
"-i",
|
||||
help="input image filename",
|
||||
type=click.Path(exists=True, dir_okay=False),
|
||||
)
|
||||
@click.option(
|
||||
"--out",
|
||||
"-o",
|
||||
help="directory for output PAGE-XML files",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
required=True,
|
||||
)
|
||||
@click.option(
|
||||
"--overwrite",
|
||||
"-O",
|
||||
help="overwrite (instead of skipping) if output xml exists",
|
||||
is_flag=True,
|
||||
)
|
||||
@click.option(
|
||||
"--dir_in",
|
||||
"-di",
|
||||
help="directory of input images (instead of --image)",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
)
|
||||
@click.option(
|
||||
"--save_images",
|
||||
"-si",
|
||||
help="if a directory is given, cropped images of pages will be saved there",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
)
|
||||
@click.option(
|
||||
"--save_layout",
|
||||
"-sl",
|
||||
help="if a directory is given, plots of layout detection will be saved there",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
)
|
||||
@click.option(
|
||||
"--save_deskewed",
|
||||
"-sd",
|
||||
help="if a directory is given, plots of page deskewing will be saved there",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
)
|
||||
@click.option(
|
||||
"--save_all",
|
||||
"-sa",
|
||||
help="if a directory is given, all plots needed will be saved there",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
)
|
||||
@click.option(
|
||||
"--save_page",
|
||||
"-sp",
|
||||
help="if a directory is given, plots of page cropping will be saved there",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
)
|
||||
@click.option(
|
||||
"--enable-plotting",
|
||||
"-ep",
|
||||
is_flag=True,
|
||||
help="plot intermediary diagnostic images to files",
|
||||
)
|
||||
@click.option(
|
||||
"--allow-enhancement",
|
||||
"-ae",
|
||||
is_flag=True,
|
||||
help="check whether input image need resizing and enhancement. If so, output of resized and enhanced image and corresponding layout data will be written in out directory",
|
||||
)
|
||||
@click.option(
|
||||
"--curved-line",
|
||||
"-cl",
|
||||
is_flag=True,
|
||||
help="try to return most precise textline contours by deskewing and detecting textlines for all text regions individually. Requires much more computation.",
|
||||
)
|
||||
@click.option(
|
||||
"--full-layout",
|
||||
"-fl",
|
||||
is_flag=True,
|
||||
help="return all elements of layout, including headings and drop-capitals",
|
||||
)
|
||||
@click.option(
|
||||
"--tables",
|
||||
"-tab",
|
||||
is_flag=True,
|
||||
help="try to detect table regions",
|
||||
)
|
||||
@click.option(
|
||||
"--right2left",
|
||||
"-r2l",
|
||||
is_flag=True,
|
||||
help="extract right-to-left reading order (instead of left-to-right)",
|
||||
)
|
||||
@click.option(
|
||||
"--input_binary",
|
||||
"-ib",
|
||||
is_flag=True,
|
||||
help="In general, eynollah uses RGB as input, but if the input document is very dark, very bright or for any other reason you can turn on internal binarization here. When set, eynollah will binarize the RGB input document first.",
|
||||
)
|
||||
@click.option(
|
||||
"--allow_scaling",
|
||||
"-as",
|
||||
is_flag=True,
|
||||
help="check the scale and if needed it will scale it to perform better layout detection",
|
||||
)
|
||||
@click.option(
|
||||
"--headers_off",
|
||||
"-ho",
|
||||
is_flag=True,
|
||||
help="ignore headers role in reading order",
|
||||
)
|
||||
@click.option(
|
||||
"--ignore_page_extraction",
|
||||
"-ipe",
|
||||
is_flag=True,
|
||||
help="ignore page extraction (cropping via page frame detection model)",
|
||||
)
|
||||
@click.option(
|
||||
"--reading_order_machine_based",
|
||||
"-romb",
|
||||
is_flag=True,
|
||||
help="apply model based reading order detection",
|
||||
)
|
||||
@click.option(
|
||||
"--num_col_upper",
|
||||
"-ncu",
|
||||
default=0,
|
||||
type=click.IntRange(min=0),
|
||||
help="lower limit of columns in document image; 0 means autodetected from model",
|
||||
)
|
||||
@click.option(
|
||||
"--num_col_lower",
|
||||
"-ncl",
|
||||
default=0,
|
||||
type=click.IntRange(min=0),
|
||||
help="upper limit of columns in document image; 0 means autodetected from model",
|
||||
)
|
||||
@click.option(
|
||||
"--threshold_art_class_layout",
|
||||
"-tharl",
|
||||
default=0.1,
|
||||
type=click.FloatRange(min=0.0, max=1.0),
|
||||
help="confidence threshold of artifical boundary class during region detection",
|
||||
)
|
||||
@click.option(
|
||||
"--threshold_art_class_textline",
|
||||
"-thart",
|
||||
default=0.1,
|
||||
type=click.FloatRange(min=0.0, max=1.0),
|
||||
help="confidence threshold of artifical boundary class during textline detection",
|
||||
)
|
||||
@click.option(
|
||||
"--skip_layout_and_reading_order",
|
||||
"-slro",
|
||||
is_flag=True,
|
||||
help="ignore layout detection and reading order, i.e. textline detection will be done within entire printspace, and textline contours will be written into a single overall text region.",
|
||||
)
|
||||
@click.option(
|
||||
"--num-jobs",
|
||||
"-j",
|
||||
default=0,
|
||||
type=click.IntRange(min=0),
|
||||
help="number of parallel images to process (for --dir_in mode; also helps better utilise GPU if available); 0 means based on autodetected number of processor cores",
|
||||
)
|
||||
@click.option(
|
||||
"--halt-fail",
|
||||
"-H",
|
||||
default=0,
|
||||
type=click.FloatRange(min=0),
|
||||
help="abort when number of failed images exceeds this value (if >=1) or ratio of failed over total images exceeds this value (if <1); 0 means ignore failures",
|
||||
)
|
||||
@click.option(
|
||||
"--device",
|
||||
"-D",
|
||||
help="placement of computations in predictors for each model type; if none (by default), will try to use first available GPU or fall back to CPU; set string to force using a device (e.g. 'GPU0', 'GPU1' or 'CPU'). Can also be a comma-separated list of model category to device mappings (e.g. 'col_classifier:CPU,page:GPU0,*:GPU1')",
|
||||
)
|
||||
@click.pass_context
|
||||
def layout_cli(
|
||||
ctx,
|
||||
image,
|
||||
out,
|
||||
overwrite,
|
||||
dir_in,
|
||||
save_images,
|
||||
save_layout,
|
||||
save_deskewed,
|
||||
save_all,
|
||||
save_page,
|
||||
enable_plotting,
|
||||
allow_enhancement,
|
||||
curved_line,
|
||||
full_layout,
|
||||
tables,
|
||||
right2left,
|
||||
input_binary,
|
||||
allow_scaling,
|
||||
headers_off,
|
||||
reading_order_machine_based,
|
||||
num_col_upper,
|
||||
num_col_lower,
|
||||
threshold_art_class_textline,
|
||||
threshold_art_class_layout,
|
||||
skip_layout_and_reading_order,
|
||||
ignore_page_extraction,
|
||||
num_jobs,
|
||||
halt_fail,
|
||||
device,
|
||||
):
|
||||
"""
|
||||
Detect Layout (with optional image enhancement and reading order detection)
|
||||
"""
|
||||
from ..eynollah import Eynollah
|
||||
assert enable_plotting or not save_layout, "Plotting with -sl also requires -ep"
|
||||
assert enable_plotting or not save_deskewed, "Plotting with -sd also requires -ep"
|
||||
assert enable_plotting or not save_all, "Plotting with -sa also requires -ep"
|
||||
assert enable_plotting or not save_page, "Plotting with -sp also requires -ep"
|
||||
assert enable_plotting or not save_images, "Plotting with -si also requires -ep"
|
||||
assert not enable_plotting or save_layout or save_deskewed or save_all or save_page or save_images or allow_enhancement, \
|
||||
"Plotting with -ep also requires -sl, -sd, -sa, -sp, -si or -ae"
|
||||
assert bool(image) != bool(dir_in), "Either -i (single input) or -di (directory) must be provided, but not both."
|
||||
eynollah = Eynollah(
|
||||
model_zoo=ctx.obj.model_zoo,
|
||||
device=device,
|
||||
enable_plotting=enable_plotting,
|
||||
allow_enhancement=allow_enhancement,
|
||||
curved_line=curved_line,
|
||||
full_layout=full_layout,
|
||||
tables=tables,
|
||||
right2left=right2left,
|
||||
input_binary=input_binary,
|
||||
allow_scaling=allow_scaling,
|
||||
headers_off=headers_off,
|
||||
ignore_page_extraction=ignore_page_extraction,
|
||||
reading_order_machine_based=reading_order_machine_based,
|
||||
num_col_upper=num_col_upper,
|
||||
num_col_lower=num_col_lower,
|
||||
skip_layout_and_reading_order=skip_layout_and_reading_order,
|
||||
threshold_art_class_textline=threshold_art_class_textline,
|
||||
threshold_art_class_layout=threshold_art_class_layout,
|
||||
)
|
||||
eynollah.run(overwrite=overwrite,
|
||||
image_filename=image,
|
||||
dir_in=dir_in,
|
||||
dir_out=out,
|
||||
dir_of_cropped_images=save_images,
|
||||
dir_of_layout=save_layout,
|
||||
dir_of_deskewed=save_deskewed,
|
||||
dir_of_all=save_all,
|
||||
dir_save_page=save_page,
|
||||
num_jobs=num_jobs,
|
||||
halt_fail=halt_fail,
|
||||
)
|
||||
|
||||
|
|
@ -1,69 +0,0 @@
|
|||
from pathlib import Path
|
||||
from typing import Set, Tuple
|
||||
import click
|
||||
|
||||
from eynollah.model_zoo.default_specs import MODELS_VERSION
|
||||
|
||||
@click.group()
|
||||
@click.pass_context
|
||||
def models_cli(
|
||||
ctx,
|
||||
):
|
||||
"""
|
||||
Organize models for the various runners in eynollah.
|
||||
"""
|
||||
assert ctx.obj.model_zoo
|
||||
|
||||
|
||||
@models_cli.command('list')
|
||||
@click.pass_context
|
||||
def list_models(
|
||||
ctx,
|
||||
):
|
||||
"""
|
||||
List all the models in the zoo
|
||||
"""
|
||||
print(f"Model basedir: {ctx.obj.model_zoo.model_basedir}")
|
||||
print(f"Model overrides: {ctx.obj.model_zoo.model_overrides}")
|
||||
print(ctx.obj.model_zoo)
|
||||
|
||||
|
||||
@models_cli.command('package')
|
||||
@click.option(
|
||||
'--set-version', '-V', 'version', help="Version to use for packaging", default=MODELS_VERSION, show_default=True
|
||||
)
|
||||
@click.argument('output_dir')
|
||||
@click.pass_context
|
||||
def package(
|
||||
ctx,
|
||||
version,
|
||||
output_dir,
|
||||
):
|
||||
"""
|
||||
Generate shell code to copy all the models in the zoo into properly named folders in OUTPUT_DIR for distribution.
|
||||
|
||||
eynollah models -m SRC package OUTPUT_DIR
|
||||
|
||||
SRC should contain a directory "models_eynollah" containing all the models.
|
||||
"""
|
||||
mkdirs: Set[Path] = set([])
|
||||
copies: Set[Tuple[Path, Path]] = set([])
|
||||
for spec in ctx.obj.model_zoo.specs.specs:
|
||||
# skip these as they are dependent on the ocr model
|
||||
if spec.category in ('num_to_char', 'characters'):
|
||||
continue
|
||||
src: Path = ctx.obj.model_zoo.model_path(spec.category, spec.variant)
|
||||
# Only copy the top-most directory relative to models_eynollah
|
||||
while src.parent.name != 'models_eynollah':
|
||||
src = src.parent
|
||||
for dist in spec.dists:
|
||||
dist_dir = Path(f"{output_dir}/models_{dist}_{version}/models_eynollah")
|
||||
copies.add((src, dist_dir))
|
||||
mkdirs.add(dist_dir)
|
||||
for dir in mkdirs:
|
||||
print(f"mkdir -vp {dir}")
|
||||
for (src, dst) in copies:
|
||||
print(f"cp -vr {src} {dst}")
|
||||
for dir in mkdirs:
|
||||
zip_path = Path(f'../{dir.parent.name}.zip')
|
||||
print(f"(cd {dir}/..; zip -vr {zip_path} models_eynollah)")
|
||||
|
|
@ -1,103 +0,0 @@
|
|||
import click
|
||||
|
||||
@click.command()
|
||||
@click.option(
|
||||
"--image",
|
||||
"-i",
|
||||
help="input image filename",
|
||||
type=click.Path(exists=True, dir_okay=False),
|
||||
)
|
||||
@click.option(
|
||||
"--dir_in",
|
||||
"-di",
|
||||
help="directory of input images (instead of --image)",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
)
|
||||
@click.option(
|
||||
"--dir_in_bin",
|
||||
"-dib",
|
||||
help=("directory of binarized images (in addition to --dir_in for RGB images; filename stems must match the RGB image files, with '.png' \n Perform prediction using both RGB and binary images. (This does not necessarily improve results, however it may be beneficial for certain document images."),
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
)
|
||||
@click.option(
|
||||
"--dir_xmls",
|
||||
"-dx",
|
||||
help="directory of input PAGE-XML files (in addition to --dir_in; filename stems must match the image files, with '.xml' suffix).",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
required=True,
|
||||
)
|
||||
@click.option(
|
||||
"--out",
|
||||
"-o",
|
||||
help="directory for output PAGE-XML files",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
required=True,
|
||||
)
|
||||
@click.option(
|
||||
"--dir_out_image_text",
|
||||
"-doit",
|
||||
help="directory for output images, newly rendered with predicted text",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
)
|
||||
@click.option(
|
||||
"--overwrite",
|
||||
"-O",
|
||||
help="overwrite (instead of skipping) if output xml exists",
|
||||
is_flag=True,
|
||||
)
|
||||
@click.option(
|
||||
"--tr_ocr",
|
||||
"-trocr/-notrocr",
|
||||
is_flag=True,
|
||||
help="if this parameter set to true, transformer ocr will be applied, otherwise cnn_rnn model.",
|
||||
)
|
||||
@click.option(
|
||||
"--do_not_mask_with_textline_contour",
|
||||
"-nmtc/-mtc",
|
||||
is_flag=True,
|
||||
help="if this parameter set to true, cropped textline images will not be masked with textline contour.",
|
||||
)
|
||||
@click.option(
|
||||
"--batch_size",
|
||||
"-bs",
|
||||
help="number of inference batch size. Default b_s for trocr and cnn_rnn models are 2 and 8 respectively",
|
||||
)
|
||||
@click.option(
|
||||
"--min_conf_value_of_textline_text",
|
||||
"-min_conf",
|
||||
help="minimum OCR confidence value. Text lines with a confidence value lower than this threshold will not be included in the output XML file.",
|
||||
)
|
||||
@click.pass_context
|
||||
def ocr_cli(
|
||||
ctx,
|
||||
image,
|
||||
dir_in,
|
||||
dir_in_bin,
|
||||
dir_xmls,
|
||||
out,
|
||||
dir_out_image_text,
|
||||
overwrite,
|
||||
tr_ocr,
|
||||
do_not_mask_with_textline_contour,
|
||||
batch_size,
|
||||
min_conf_value_of_textline_text,
|
||||
):
|
||||
"""
|
||||
Recognize text with a CNN/RNN or transformer ML model.
|
||||
"""
|
||||
assert bool(image) ^ bool(dir_in), "Either -i (single image) or -di (directory) must be provided, but not both."
|
||||
from ..eynollah_ocr import Eynollah_ocr
|
||||
eynollah_ocr = Eynollah_ocr(
|
||||
model_zoo=ctx.obj.model_zoo,
|
||||
tr_ocr=tr_ocr,
|
||||
do_not_mask_with_textline_contour=do_not_mask_with_textline_contour,
|
||||
batch_size=batch_size,
|
||||
min_conf_value_of_textline_text=min_conf_value_of_textline_text)
|
||||
eynollah_ocr.run(overwrite=overwrite,
|
||||
dir_in=dir_in,
|
||||
dir_in_bin=dir_in_bin,
|
||||
image_filename=image,
|
||||
dir_xmls=dir_xmls,
|
||||
dir_out_image_text=dir_out_image_text,
|
||||
dir_out=out,
|
||||
)
|
||||
|
|
@ -1,35 +0,0 @@
|
|||
import click
|
||||
|
||||
@click.command()
|
||||
@click.option(
|
||||
"--input",
|
||||
"-i",
|
||||
help="PAGE-XML input filename",
|
||||
type=click.Path(exists=True, dir_okay=False),
|
||||
)
|
||||
@click.option(
|
||||
"--dir_in",
|
||||
"-di",
|
||||
help="directory of PAGE-XML input files (instead of --input)",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
)
|
||||
@click.option(
|
||||
"--out",
|
||||
"-o",
|
||||
help="directory for output images",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
required=True,
|
||||
)
|
||||
@click.pass_context
|
||||
def readingorder_cli(ctx, input, dir_in, out):
|
||||
"""
|
||||
Generate ReadingOrder with a ML model
|
||||
"""
|
||||
from ..mb_ro_on_layout import machine_based_reading_order_on_layout
|
||||
assert bool(input) != bool(dir_in), "Either -i (single input) or -di (directory) must be provided, but not both."
|
||||
orderer = machine_based_reading_order_on_layout(model_zoo=ctx.obj.model_zoo)
|
||||
orderer.run(xml_filename=input,
|
||||
dir_in=dir_in,
|
||||
dir_out=out,
|
||||
)
|
||||
|
||||
|
|
@ -1,281 +0,0 @@
|
|||
"""
|
||||
extract images?
|
||||
"""
|
||||
|
||||
from concurrent.futures import ProcessPoolExecutor
|
||||
import logging
|
||||
from multiprocessing import cpu_count
|
||||
import os
|
||||
import time
|
||||
from typing import Optional
|
||||
from pathlib import Path
|
||||
import tensorflow as tf
|
||||
import numpy as np
|
||||
import cv2
|
||||
|
||||
from eynollah.utils.contour import filter_contours_area_of_image, return_contours_of_image, return_contours_of_interested_region
|
||||
from eynollah.utils.resize import resize_image
|
||||
|
||||
from .model_zoo.model_zoo import EynollahModelZoo
|
||||
from .eynollah import Eynollah
|
||||
from .utils import box2rect, is_image_filename
|
||||
from .plot import EynollahPlotter
|
||||
|
||||
class EynollahImageExtractor(Eynollah):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
*,
|
||||
model_zoo: EynollahModelZoo,
|
||||
enable_plotting : bool = False,
|
||||
input_binary : bool = False,
|
||||
ignore_page_extraction : bool = False,
|
||||
num_col_upper : Optional[int] = None,
|
||||
num_col_lower : Optional[int] = None,
|
||||
full_layout : bool = False,
|
||||
tables : bool = False,
|
||||
curved_line : bool = False,
|
||||
allow_enhancement : bool = False,
|
||||
|
||||
):
|
||||
self.logger = logging.getLogger('eynollah.extract_images')
|
||||
self.model_zoo = model_zoo
|
||||
self.plotter = None
|
||||
self.tables = tables
|
||||
self.curved_line = curved_line
|
||||
self.allow_enhancement = allow_enhancement
|
||||
|
||||
self.enable_plotting = enable_plotting
|
||||
# --input-binary sensible if image is very dark, if layout is not working.
|
||||
self.input_binary = input_binary
|
||||
self.ignore_page_extraction = ignore_page_extraction
|
||||
self.full_layout = full_layout
|
||||
if num_col_upper:
|
||||
self.num_col_upper = int(num_col_upper)
|
||||
else:
|
||||
self.num_col_upper = num_col_upper
|
||||
if num_col_lower:
|
||||
self.num_col_lower = int(num_col_lower)
|
||||
else:
|
||||
self.num_col_lower = num_col_lower
|
||||
|
||||
# for parallelization of CPU-intensive tasks:
|
||||
self.executor = ProcessPoolExecutor(max_workers=cpu_count())
|
||||
|
||||
t_start = time.time()
|
||||
|
||||
try:
|
||||
for device in tf.config.list_physical_devices('GPU'):
|
||||
tf.config.experimental.set_memory_growth(device, True)
|
||||
except:
|
||||
self.logger.warning("no GPU device available")
|
||||
|
||||
self.logger.info("Loading models...")
|
||||
self.setup_models()
|
||||
self.logger.info(f"Model initialization complete ({time.time() - t_start:.1f}s)")
|
||||
|
||||
def setup_models(self):
|
||||
|
||||
loadable = [
|
||||
"col_classifier",
|
||||
"binarization",
|
||||
"page",
|
||||
"extract_images",
|
||||
]
|
||||
self.model_zoo.load_models(*loadable)
|
||||
|
||||
def get_regions_light_v_extract_only_images(self,img, num_col_classifier):
|
||||
self.logger.debug("enter get_regions_extract_images_only")
|
||||
erosion_hurts = False
|
||||
img_org = np.copy(img)
|
||||
img_height_h = img_org.shape[0]
|
||||
img_width_h = img_org.shape[1]
|
||||
|
||||
if num_col_classifier == 1:
|
||||
img_w_new = 700
|
||||
elif num_col_classifier == 2:
|
||||
img_w_new = 900
|
||||
elif num_col_classifier == 3:
|
||||
img_w_new = 1500
|
||||
elif num_col_classifier == 4:
|
||||
img_w_new = 1800
|
||||
elif num_col_classifier == 5:
|
||||
img_w_new = 2200
|
||||
elif num_col_classifier == 6:
|
||||
img_w_new = 2500
|
||||
else:
|
||||
raise ValueError("num_col_classifier must be in range 1..6")
|
||||
img_h_new = int(img.shape[0] / float(img.shape[1]) * img_w_new)
|
||||
img_resized = resize_image(img,img_h_new, img_w_new )
|
||||
|
||||
prediction_regions_org, _ = self.do_prediction_new_concept(True, img_resized, self.model_zoo.get("extract_images"))
|
||||
|
||||
prediction_regions_org = resize_image(prediction_regions_org,img_height_h, img_width_h )
|
||||
image_page, page_coord, cont_page = self.extract_page()
|
||||
|
||||
prediction_regions_org = prediction_regions_org[page_coord[0] : page_coord[1], page_coord[2] : page_coord[3]]
|
||||
prediction_regions_org=prediction_regions_org[:,:,0]
|
||||
|
||||
mask_seps_only = (prediction_regions_org[:,:] ==3)*1
|
||||
mask_texts_only = (prediction_regions_org[:,:] ==1)*1
|
||||
mask_images_only=(prediction_regions_org[:,:] ==2)*1
|
||||
|
||||
polygons_seplines, hir_seplines = return_contours_of_image(mask_seps_only)
|
||||
polygons_seplines = filter_contours_area_of_image(
|
||||
mask_seps_only, polygons_seplines, hir_seplines, max_area=1, min_area=0.00001, dilate=1)
|
||||
|
||||
polygons_of_only_texts = return_contours_of_interested_region(mask_texts_only,1,0.00001)
|
||||
polygons_of_only_seps = return_contours_of_interested_region(mask_seps_only,1,0.00001)
|
||||
|
||||
text_regions_p_true = np.zeros(prediction_regions_org.shape)
|
||||
text_regions_p_true = cv2.fillPoly(text_regions_p_true, pts = polygons_of_only_seps, color=(3,3,3))
|
||||
|
||||
text_regions_p_true[:,:][mask_images_only[:,:] == 1] = 2
|
||||
text_regions_p_true = cv2.fillPoly(text_regions_p_true, pts=polygons_of_only_texts, color=(1,1,1))
|
||||
|
||||
text_regions_p_true[text_regions_p_true.shape[0]-15:text_regions_p_true.shape[0], :] = 0
|
||||
text_regions_p_true[:, text_regions_p_true.shape[1]-15:text_regions_p_true.shape[1]] = 0
|
||||
|
||||
##polygons_of_images = return_contours_of_interested_region(text_regions_p_true, 2, 0.0001)
|
||||
polygons_of_images = return_contours_of_interested_region(text_regions_p_true, 2, 0.001)
|
||||
|
||||
polygons_of_images_fin = []
|
||||
for ploy_img_ind in polygons_of_images:
|
||||
box = _, _, w, h = cv2.boundingRect(ploy_img_ind)
|
||||
if h < 150 or w < 150:
|
||||
pass
|
||||
else:
|
||||
page_coord_img = box2rect(box) # type: ignore
|
||||
polygons_of_images_fin.append(np.array([[page_coord_img[2], page_coord_img[0]],
|
||||
[page_coord_img[3], page_coord_img[0]],
|
||||
[page_coord_img[3], page_coord_img[1]],
|
||||
[page_coord_img[2], page_coord_img[1]]]))
|
||||
|
||||
self.logger.debug("exit get_regions_extract_images_only")
|
||||
return (text_regions_p_true,
|
||||
erosion_hurts,
|
||||
polygons_seplines,
|
||||
polygons_of_images_fin,
|
||||
image_page,
|
||||
page_coord,
|
||||
cont_page)
|
||||
|
||||
def run(self,
|
||||
overwrite: bool = False,
|
||||
image_filename: Optional[str] = None,
|
||||
dir_in: Optional[str] = None,
|
||||
dir_out: Optional[str] = None,
|
||||
dir_of_cropped_images: Optional[str] = None,
|
||||
dir_of_layout: Optional[str] = None,
|
||||
dir_of_deskewed: Optional[str] = None,
|
||||
dir_of_all: Optional[str] = None,
|
||||
dir_save_page: Optional[str] = None,
|
||||
):
|
||||
"""
|
||||
Get image and scales, then extract the page of scanned image
|
||||
"""
|
||||
self.logger.debug("enter run")
|
||||
t0_tot = time.time()
|
||||
# Log enabled features directly
|
||||
enabled_modes = []
|
||||
if self.full_layout:
|
||||
enabled_modes.append("Full layout analysis")
|
||||
if self.tables:
|
||||
enabled_modes.append("Table detection")
|
||||
if enabled_modes:
|
||||
self.logger.info("Enabled modes: " + ", ".join(enabled_modes))
|
||||
if self.enable_plotting:
|
||||
self.logger.info("Saving debug plots")
|
||||
if dir_of_cropped_images:
|
||||
self.logger.info(f"Saving cropped images to: {dir_of_cropped_images}")
|
||||
if dir_of_layout:
|
||||
self.logger.info(f"Saving layout plots to: {dir_of_layout}")
|
||||
if dir_of_deskewed:
|
||||
self.logger.info(f"Saving deskewed images to: {dir_of_deskewed}")
|
||||
|
||||
if dir_in:
|
||||
ls_imgs = [os.path.join(dir_in, image_filename)
|
||||
for image_filename in filter(is_image_filename,
|
||||
os.listdir(dir_in))]
|
||||
elif image_filename:
|
||||
ls_imgs = [image_filename]
|
||||
else:
|
||||
raise ValueError("run requires either a single image filename or a directory")
|
||||
|
||||
for img_filename in ls_imgs:
|
||||
self.logger.info(img_filename)
|
||||
t0 = time.time()
|
||||
|
||||
self.reset_file_name_dir(img_filename, dir_out)
|
||||
if self.enable_plotting:
|
||||
self.plotter = EynollahPlotter(dir_out=dir_out,
|
||||
dir_of_all=dir_of_all,
|
||||
dir_save_page=dir_save_page,
|
||||
dir_of_deskewed=dir_of_deskewed,
|
||||
dir_of_cropped_images=dir_of_cropped_images,
|
||||
dir_of_layout=dir_of_layout,
|
||||
image_filename_stem=Path(img_filename).stem)
|
||||
#print("text region early -11 in %.1fs", time.time() - t0)
|
||||
if os.path.exists(self.writer.output_filename):
|
||||
if overwrite:
|
||||
self.logger.warning("will overwrite existing output file '%s'", self.writer.output_filename)
|
||||
else:
|
||||
self.logger.warning("will skip input for existing output file '%s'", self.writer.output_filename)
|
||||
continue
|
||||
|
||||
pcgts = self.run_single()
|
||||
self.logger.info("Job done in %.1fs", time.time() - t0)
|
||||
self.writer.write_pagexml(pcgts)
|
||||
|
||||
if dir_in:
|
||||
self.logger.info("All jobs done in %.1fs", time.time() - t0_tot)
|
||||
|
||||
def run_single(self):
|
||||
t0 = time.time()
|
||||
|
||||
self.logger.info(f"Processing file: {self.writer.image_filename}")
|
||||
self.logger.info("Step 1/5: Image Enhancement")
|
||||
|
||||
img_res, is_image_enhanced, num_col_classifier, _ = \
|
||||
self.run_enhancement()
|
||||
|
||||
self.logger.info(f"Image: {self.image.shape[1]}x{self.image.shape[0]}, "
|
||||
f"{self.dpi} DPI, {num_col_classifier} columns")
|
||||
if is_image_enhanced:
|
||||
self.logger.info("Enhancement applied")
|
||||
|
||||
self.logger.info(f"Enhancement complete ({time.time() - t0:.1f}s)")
|
||||
|
||||
|
||||
# Image Extraction Mode
|
||||
self.logger.info("Step 2/5: Image Extraction Mode")
|
||||
|
||||
_, _, _, polygons_of_images, \
|
||||
image_page, page_coord, cont_page = \
|
||||
self.get_regions_light_v_extract_only_images(img_res, num_col_classifier)
|
||||
|
||||
pcgts = self.writer.build_pagexml_no_full_layout(
|
||||
found_polygons_text_region=[],
|
||||
page_coord=page_coord,
|
||||
order_of_texts=[],
|
||||
all_found_textline_polygons=[],
|
||||
all_box_coord=[],
|
||||
found_polygons_text_region_img=polygons_of_images,
|
||||
found_polygons_marginals_left=[],
|
||||
found_polygons_marginals_right=[],
|
||||
all_found_textline_polygons_marginals_left=[],
|
||||
all_found_textline_polygons_marginals_right=[],
|
||||
all_box_coord_marginals_left=[],
|
||||
all_box_coord_marginals_right=[],
|
||||
slopes=[],
|
||||
slopes_marginals_left=[],
|
||||
slopes_marginals_right=[],
|
||||
cont_page=cont_page,
|
||||
polygons_seplines=[],
|
||||
found_polygons_tables=[],
|
||||
)
|
||||
if self.plotter:
|
||||
self.plotter.write_images_into_directory(polygons_of_images, image_page)
|
||||
|
||||
self.logger.info("Image extraction complete")
|
||||
return pcgts
|
||||
File diff suppressed because it is too large
Load diff
|
|
@ -1,13 +0,0 @@
|
|||
"""
|
||||
Load libraries with possible race conditions once. This must be imported as the first module of eynollah.
|
||||
"""
|
||||
import os
|
||||
os.environ['TF_USE_LEGACY_KERAS'] = '1' # avoid Keras 3 after TF 2.15
|
||||
|
||||
from ocrd_utils import tf_disable_interactive_logs
|
||||
from torch import *
|
||||
tf_disable_interactive_logs()
|
||||
import tensorflow.keras
|
||||
from shapely import *
|
||||
imported_libs = True
|
||||
__all__ = ['imported_libs']
|
||||
|
|
@ -1,837 +0,0 @@
|
|||
# FIXME: fix all of those...
|
||||
# pyright: reportOptionalSubscript=false
|
||||
|
||||
from logging import Logger, getLogger
|
||||
from typing import List, Optional
|
||||
from pathlib import Path
|
||||
import os
|
||||
import gc
|
||||
import math
|
||||
from dataclasses import dataclass
|
||||
|
||||
import cv2
|
||||
from cv2.typing import MatLike
|
||||
from xml.etree import ElementTree as ET
|
||||
from PIL import Image, ImageDraw
|
||||
import numpy as np
|
||||
from eynollah.model_zoo import EynollahModelZoo
|
||||
from eynollah.utils.font import get_font
|
||||
from eynollah.utils.xml import etree_namespace_for_element_tag
|
||||
try:
|
||||
import torch
|
||||
except ImportError:
|
||||
torch = None
|
||||
|
||||
|
||||
from .utils import is_image_filename
|
||||
from .utils.resize import resize_image
|
||||
from .utils.utils_ocr import (
|
||||
break_curved_line_into_small_pieces_and_then_merge,
|
||||
decode_batch_predictions,
|
||||
fit_text_single_line,
|
||||
get_contours_and_bounding_boxes,
|
||||
get_orientation_moments,
|
||||
preprocess_and_resize_image_for_ocrcnn_model,
|
||||
return_textlines_split_if_needed,
|
||||
rotate_image_with_padding,
|
||||
)
|
||||
|
||||
# TODO: refine typing
|
||||
@dataclass
|
||||
class EynollahOcrResult:
|
||||
extracted_texts_merged: List
|
||||
extracted_conf_value_merged: Optional[List]
|
||||
cropped_lines_region_indexer: List
|
||||
total_bb_coordinates:List
|
||||
|
||||
class Eynollah_ocr:
|
||||
def __init__(
|
||||
self,
|
||||
*,
|
||||
model_zoo: EynollahModelZoo,
|
||||
tr_ocr=False,
|
||||
batch_size: Optional[int]=None,
|
||||
do_not_mask_with_textline_contour: bool=False,
|
||||
min_conf_value_of_textline_text : Optional[float]=None,
|
||||
logger: Optional[Logger]=None,
|
||||
):
|
||||
self.tr_ocr = tr_ocr
|
||||
# masking for OCR and GT generation, relevant for skewed lines and bounding boxes
|
||||
self.do_not_mask_with_textline_contour = do_not_mask_with_textline_contour
|
||||
self.logger = logger if logger else getLogger('eynollah.ocr')
|
||||
self.model_zoo = model_zoo
|
||||
|
||||
self.min_conf_value_of_textline_text = min_conf_value_of_textline_text if min_conf_value_of_textline_text else 0.3
|
||||
self.b_s = 2 if batch_size is None and tr_ocr else 8 if batch_size is None else batch_size
|
||||
|
||||
if tr_ocr:
|
||||
self.model_zoo.load_models('trocr_processor')
|
||||
self.model_zoo.load_models(['ocr', 'tr'])
|
||||
self.model_zoo.get('ocr').to(self.device)
|
||||
else:
|
||||
self.model_zoo.load_models('ocr')
|
||||
self.model_zoo.load_models('num_to_char')
|
||||
self.model_zoo.load_models('characters')
|
||||
self.end_character = len(self.model_zoo.get('characters')) + 2
|
||||
|
||||
@property
|
||||
def device(self):
|
||||
assert torch
|
||||
if torch.cuda.is_available():
|
||||
self.logger.info("Using GPU acceleration")
|
||||
return torch.device("cuda:0")
|
||||
else:
|
||||
self.logger.info("Using CPU processing")
|
||||
return torch.device("cpu")
|
||||
|
||||
def run_trocr(
|
||||
self,
|
||||
*,
|
||||
img: MatLike,
|
||||
page_tree: ET.ElementTree,
|
||||
page_ns,
|
||||
tr_ocr_input_height_and_width,
|
||||
) -> EynollahOcrResult:
|
||||
|
||||
total_bb_coordinates = []
|
||||
|
||||
|
||||
cropped_lines = []
|
||||
cropped_lines_region_indexer = []
|
||||
cropped_lines_meging_indexing = []
|
||||
|
||||
extracted_texts = []
|
||||
|
||||
indexer_text_region = 0
|
||||
indexer_b_s = 0
|
||||
|
||||
for nn in page_tree.getroot().iter(f'{{{page_ns}}}TextRegion'):
|
||||
for child_textregion in nn:
|
||||
if child_textregion.tag.endswith("TextLine"):
|
||||
|
||||
for child_textlines in child_textregion:
|
||||
if child_textlines.tag.endswith("Coords"):
|
||||
cropped_lines_region_indexer.append(indexer_text_region)
|
||||
p_h=child_textlines.attrib['points'].split(' ')
|
||||
textline_coords = np.array( [ [int(x.split(',')[0]),
|
||||
int(x.split(',')[1]) ]
|
||||
for x in p_h] )
|
||||
x,y,w,h = cv2.boundingRect(textline_coords)
|
||||
|
||||
total_bb_coordinates.append([x,y,w,h])
|
||||
|
||||
h2w_ratio = h/float(w)
|
||||
|
||||
img_poly_on_img = np.copy(img)
|
||||
mask_poly = np.zeros(img.shape)
|
||||
mask_poly = cv2.fillPoly(mask_poly, pts=[textline_coords], color=(1, 1, 1))
|
||||
|
||||
mask_poly = mask_poly[y:y+h, x:x+w, :]
|
||||
img_crop = img_poly_on_img[y:y+h, x:x+w, :]
|
||||
img_crop[mask_poly==0] = 255
|
||||
|
||||
self.logger.debug("processing %d lines for '%s'",
|
||||
len(cropped_lines), nn.attrib['id'])
|
||||
if h2w_ratio > 0.1:
|
||||
cropped_lines.append(resize_image(img_crop,
|
||||
tr_ocr_input_height_and_width,
|
||||
tr_ocr_input_height_and_width) )
|
||||
cropped_lines_meging_indexing.append(0)
|
||||
indexer_b_s+=1
|
||||
if indexer_b_s==self.b_s:
|
||||
imgs = cropped_lines[:]
|
||||
cropped_lines = []
|
||||
indexer_b_s = 0
|
||||
|
||||
pixel_values_merged = self.model_zoo.get('trocr_processor')(imgs, return_tensors="pt").pixel_values
|
||||
generated_ids_merged = self.model_zoo.get('ocr').generate(
|
||||
pixel_values_merged.to(self.device))
|
||||
generated_text_merged = self.model_zoo.get('trocr_processor').batch_decode(
|
||||
generated_ids_merged, skip_special_tokens=True)
|
||||
|
||||
extracted_texts = extracted_texts + generated_text_merged
|
||||
|
||||
else:
|
||||
splited_images, _ = return_textlines_split_if_needed(img_crop, None)
|
||||
#print(splited_images)
|
||||
if splited_images:
|
||||
cropped_lines.append(resize_image(splited_images[0],
|
||||
tr_ocr_input_height_and_width,
|
||||
tr_ocr_input_height_and_width))
|
||||
cropped_lines_meging_indexing.append(1)
|
||||
indexer_b_s+=1
|
||||
|
||||
if indexer_b_s==self.b_s:
|
||||
imgs = cropped_lines[:]
|
||||
cropped_lines = []
|
||||
indexer_b_s = 0
|
||||
|
||||
pixel_values_merged = self.model_zoo.get('trocr_processor')(imgs, return_tensors="pt").pixel_values
|
||||
generated_ids_merged = self.model_zoo.get('ocr').generate(
|
||||
pixel_values_merged.to(self.device))
|
||||
generated_text_merged = self.model_zoo.get('trocr_processor').batch_decode(
|
||||
generated_ids_merged, skip_special_tokens=True)
|
||||
|
||||
extracted_texts = extracted_texts + generated_text_merged
|
||||
|
||||
|
||||
cropped_lines.append(resize_image(splited_images[1],
|
||||
tr_ocr_input_height_and_width,
|
||||
tr_ocr_input_height_and_width))
|
||||
cropped_lines_meging_indexing.append(-1)
|
||||
indexer_b_s+=1
|
||||
|
||||
if indexer_b_s==self.b_s:
|
||||
imgs = cropped_lines[:]
|
||||
cropped_lines = []
|
||||
indexer_b_s = 0
|
||||
|
||||
pixel_values_merged = self.model_zoo.get('trocr_processor')(imgs, return_tensors="pt").pixel_values
|
||||
generated_ids_merged = self.model_zoo.get('ocr').generate(
|
||||
pixel_values_merged.to(self.device))
|
||||
generated_text_merged = self.model_zoo.get('trocr_processor').batch_decode(
|
||||
generated_ids_merged, skip_special_tokens=True)
|
||||
|
||||
extracted_texts = extracted_texts + generated_text_merged
|
||||
|
||||
else:
|
||||
cropped_lines.append(img_crop)
|
||||
cropped_lines_meging_indexing.append(0)
|
||||
indexer_b_s+=1
|
||||
|
||||
if indexer_b_s==self.b_s:
|
||||
imgs = cropped_lines[:]
|
||||
cropped_lines = []
|
||||
indexer_b_s = 0
|
||||
|
||||
pixel_values_merged = self.model_zoo.get('trocr_processor')(imgs, return_tensors="pt").pixel_values
|
||||
generated_ids_merged = self.model_zoo.get('ocr').generate(
|
||||
pixel_values_merged.to(self.device))
|
||||
generated_text_merged = self.model_zoo.get('trocr_processor').batch_decode(
|
||||
generated_ids_merged, skip_special_tokens=True)
|
||||
|
||||
extracted_texts = extracted_texts + generated_text_merged
|
||||
|
||||
|
||||
|
||||
indexer_text_region = indexer_text_region +1
|
||||
|
||||
if indexer_b_s!=0:
|
||||
imgs = cropped_lines[:]
|
||||
cropped_lines = []
|
||||
indexer_b_s = 0
|
||||
|
||||
pixel_values_merged = self.model_zoo.get('trocr_processor')(imgs, return_tensors="pt").pixel_values
|
||||
generated_ids_merged = self.model_zoo.get('ocr').generate(pixel_values_merged.to(self.device))
|
||||
generated_text_merged = self.model_zoo.get('trocr_processor').batch_decode(generated_ids_merged, skip_special_tokens=True)
|
||||
|
||||
extracted_texts = extracted_texts + generated_text_merged
|
||||
|
||||
####extracted_texts = []
|
||||
####n_iterations = math.ceil(len(cropped_lines) / self.b_s)
|
||||
|
||||
####for i in range(n_iterations):
|
||||
####if i==(n_iterations-1):
|
||||
####n_start = i*self.b_s
|
||||
####imgs = cropped_lines[n_start:]
|
||||
####else:
|
||||
####n_start = i*self.b_s
|
||||
####n_end = (i+1)*self.b_s
|
||||
####imgs = cropped_lines[n_start:n_end]
|
||||
####pixel_values_merged = self.model_zoo.get('trocr_processor')(imgs, return_tensors="pt").pixel_values
|
||||
####generated_ids_merged = self.model_ocr.generate(
|
||||
#### pixel_values_merged.to(self.device))
|
||||
####generated_text_merged = self.model_zoo.get('trocr_processor').batch_decode(
|
||||
#### generated_ids_merged, skip_special_tokens=True)
|
||||
|
||||
####extracted_texts = extracted_texts + generated_text_merged
|
||||
|
||||
del cropped_lines
|
||||
gc.collect()
|
||||
|
||||
extracted_texts_merged = [extracted_texts[ind]
|
||||
if cropped_lines_meging_indexing[ind]==0
|
||||
else extracted_texts[ind]+" "+extracted_texts[ind+1]
|
||||
if cropped_lines_meging_indexing[ind]==1
|
||||
else None
|
||||
for ind in range(len(cropped_lines_meging_indexing))]
|
||||
|
||||
extracted_texts_merged = [ind for ind in extracted_texts_merged if ind is not None]
|
||||
#print(extracted_texts_merged, len(extracted_texts_merged))
|
||||
|
||||
return EynollahOcrResult(
|
||||
extracted_texts_merged=extracted_texts_merged,
|
||||
extracted_conf_value_merged=None,
|
||||
cropped_lines_region_indexer=cropped_lines_region_indexer,
|
||||
total_bb_coordinates=total_bb_coordinates,
|
||||
)
|
||||
|
||||
def run_cnn(
|
||||
self,
|
||||
*,
|
||||
img: MatLike,
|
||||
img_bin: Optional[MatLike],
|
||||
page_tree: ET.ElementTree,
|
||||
page_ns,
|
||||
image_width,
|
||||
image_height,
|
||||
) -> EynollahOcrResult:
|
||||
|
||||
total_bb_coordinates = []
|
||||
|
||||
cropped_lines = []
|
||||
img_crop_bin = None
|
||||
imgs_bin = None
|
||||
imgs_bin_ver_flipped = None
|
||||
cropped_lines_bin = []
|
||||
cropped_lines_ver_index = []
|
||||
cropped_lines_region_indexer = []
|
||||
cropped_lines_meging_indexing = []
|
||||
|
||||
indexer_text_region = 0
|
||||
for nn in page_tree.getroot().iter(f'{{{page_ns}}}TextRegion'):
|
||||
try:
|
||||
type_textregion = nn.attrib['type']
|
||||
except:
|
||||
type_textregion = 'paragraph'
|
||||
for child_textregion in nn:
|
||||
if child_textregion.tag.endswith("TextLine"):
|
||||
for child_textlines in child_textregion:
|
||||
if child_textlines.tag.endswith("Coords"):
|
||||
cropped_lines_region_indexer.append(indexer_text_region)
|
||||
p_h=child_textlines.attrib['points'].split(' ')
|
||||
textline_coords = np.array( [ [int(x.split(',')[0]),
|
||||
int(x.split(',')[1]) ]
|
||||
for x in p_h] )
|
||||
|
||||
x,y,w,h = cv2.boundingRect(textline_coords)
|
||||
|
||||
angle_radians = math.atan2(h, w)
|
||||
# Convert to degrees
|
||||
angle_degrees = math.degrees(angle_radians)
|
||||
if type_textregion=='drop-capital':
|
||||
angle_degrees = 0
|
||||
|
||||
total_bb_coordinates.append([x,y,w,h])
|
||||
|
||||
w_scaled = w * image_height/float(h)
|
||||
|
||||
img_poly_on_img = np.copy(img)
|
||||
if img_bin:
|
||||
img_poly_on_img_bin = np.copy(img_bin)
|
||||
img_crop_bin = img_poly_on_img_bin[y:y+h, x:x+w, :]
|
||||
|
||||
mask_poly = np.zeros(img.shape)
|
||||
mask_poly = cv2.fillPoly(mask_poly, pts=[textline_coords], color=(1, 1, 1))
|
||||
|
||||
|
||||
mask_poly = mask_poly[y:y+h, x:x+w, :]
|
||||
img_crop = img_poly_on_img[y:y+h, x:x+w, :]
|
||||
|
||||
# print(file_name, angle_degrees, w*h,
|
||||
# mask_poly[:,:,0].sum(),
|
||||
# mask_poly[:,:,0].sum() /float(w*h) ,
|
||||
# 'didi')
|
||||
|
||||
if angle_degrees > 3:
|
||||
better_des_slope = get_orientation_moments(textline_coords)
|
||||
|
||||
img_crop = rotate_image_with_padding(img_crop, better_des_slope)
|
||||
if img_bin:
|
||||
img_crop_bin = rotate_image_with_padding(img_crop_bin, better_des_slope)
|
||||
|
||||
mask_poly = rotate_image_with_padding(mask_poly, better_des_slope)
|
||||
mask_poly = mask_poly.astype('uint8')
|
||||
|
||||
#new bounding box
|
||||
x_n, y_n, w_n, h_n = get_contours_and_bounding_boxes(mask_poly[:,:,0])
|
||||
|
||||
mask_poly = mask_poly[y_n:y_n+h_n, x_n:x_n+w_n, :]
|
||||
img_crop = img_crop[y_n:y_n+h_n, x_n:x_n+w_n, :]
|
||||
|
||||
if not self.do_not_mask_with_textline_contour:
|
||||
img_crop[mask_poly==0] = 255
|
||||
if img_bin:
|
||||
img_crop_bin = img_crop_bin[y_n:y_n+h_n, x_n:x_n+w_n, :]
|
||||
if not self.do_not_mask_with_textline_contour:
|
||||
img_crop_bin[mask_poly==0] = 255
|
||||
|
||||
if mask_poly[:,:,0].sum() /float(w_n*h_n) < 0.50 and w_scaled > 90:
|
||||
if img_bin:
|
||||
img_crop, img_crop_bin = \
|
||||
break_curved_line_into_small_pieces_and_then_merge(
|
||||
img_crop, mask_poly, img_crop_bin)
|
||||
else:
|
||||
img_crop, _ = \
|
||||
break_curved_line_into_small_pieces_and_then_merge(
|
||||
img_crop, mask_poly)
|
||||
|
||||
else:
|
||||
better_des_slope = 0
|
||||
if not self.do_not_mask_with_textline_contour:
|
||||
img_crop[mask_poly==0] = 255
|
||||
if img_bin:
|
||||
if not self.do_not_mask_with_textline_contour:
|
||||
img_crop_bin[mask_poly==0] = 255
|
||||
if type_textregion=='drop-capital':
|
||||
pass
|
||||
else:
|
||||
if mask_poly[:,:,0].sum() /float(w*h) < 0.50 and w_scaled > 90:
|
||||
if img_bin:
|
||||
img_crop, img_crop_bin = \
|
||||
break_curved_line_into_small_pieces_and_then_merge(
|
||||
img_crop, mask_poly, img_crop_bin)
|
||||
else:
|
||||
img_crop, _ = \
|
||||
break_curved_line_into_small_pieces_and_then_merge(
|
||||
img_crop, mask_poly)
|
||||
|
||||
if w_scaled < 750:#1.5*image_width:
|
||||
img_fin = preprocess_and_resize_image_for_ocrcnn_model(
|
||||
img_crop, image_height, image_width)
|
||||
cropped_lines.append(img_fin)
|
||||
if abs(better_des_slope) > 45:
|
||||
cropped_lines_ver_index.append(1)
|
||||
else:
|
||||
cropped_lines_ver_index.append(0)
|
||||
|
||||
cropped_lines_meging_indexing.append(0)
|
||||
if img_bin:
|
||||
img_fin = preprocess_and_resize_image_for_ocrcnn_model(
|
||||
img_crop_bin, image_height, image_width)
|
||||
cropped_lines_bin.append(img_fin)
|
||||
else:
|
||||
splited_images, splited_images_bin = return_textlines_split_if_needed(
|
||||
img_crop, img_crop_bin if img_bin else None)
|
||||
if splited_images:
|
||||
img_fin = preprocess_and_resize_image_for_ocrcnn_model(
|
||||
splited_images[0], image_height, image_width)
|
||||
cropped_lines.append(img_fin)
|
||||
cropped_lines_meging_indexing.append(1)
|
||||
|
||||
if abs(better_des_slope) > 45:
|
||||
cropped_lines_ver_index.append(1)
|
||||
else:
|
||||
cropped_lines_ver_index.append(0)
|
||||
|
||||
img_fin = preprocess_and_resize_image_for_ocrcnn_model(
|
||||
splited_images[1], image_height, image_width)
|
||||
|
||||
cropped_lines.append(img_fin)
|
||||
cropped_lines_meging_indexing.append(-1)
|
||||
|
||||
if abs(better_des_slope) > 45:
|
||||
cropped_lines_ver_index.append(1)
|
||||
else:
|
||||
cropped_lines_ver_index.append(0)
|
||||
|
||||
if img_bin:
|
||||
img_fin = preprocess_and_resize_image_for_ocrcnn_model(
|
||||
splited_images_bin[0], image_height, image_width)
|
||||
cropped_lines_bin.append(img_fin)
|
||||
img_fin = preprocess_and_resize_image_for_ocrcnn_model(
|
||||
splited_images_bin[1], image_height, image_width)
|
||||
cropped_lines_bin.append(img_fin)
|
||||
|
||||
else:
|
||||
img_fin = preprocess_and_resize_image_for_ocrcnn_model(
|
||||
img_crop, image_height, image_width)
|
||||
cropped_lines.append(img_fin)
|
||||
cropped_lines_meging_indexing.append(0)
|
||||
|
||||
if abs(better_des_slope) > 45:
|
||||
cropped_lines_ver_index.append(1)
|
||||
else:
|
||||
cropped_lines_ver_index.append(0)
|
||||
|
||||
if img_bin:
|
||||
img_fin = preprocess_and_resize_image_for_ocrcnn_model(
|
||||
img_crop_bin, image_height, image_width)
|
||||
cropped_lines_bin.append(img_fin)
|
||||
|
||||
|
||||
indexer_text_region = indexer_text_region +1
|
||||
|
||||
extracted_texts = []
|
||||
extracted_conf_value = []
|
||||
|
||||
n_iterations = math.ceil(len(cropped_lines) / self.b_s)
|
||||
|
||||
# FIXME: copy pasta
|
||||
for i in range(n_iterations):
|
||||
if i==(n_iterations-1):
|
||||
n_start = i*self.b_s
|
||||
imgs = cropped_lines[n_start:]
|
||||
imgs = np.array(imgs)
|
||||
imgs = imgs.reshape(imgs.shape[0], image_height, image_width, 3)
|
||||
|
||||
ver_imgs = np.array( cropped_lines_ver_index[n_start:] )
|
||||
indices_ver = np.where(ver_imgs == 1)[0]
|
||||
|
||||
#print(indices_ver, 'indices_ver')
|
||||
if len(indices_ver)>0:
|
||||
imgs_ver_flipped = imgs[indices_ver, : ,: ,:]
|
||||
imgs_ver_flipped = imgs_ver_flipped[:,::-1,::-1,:]
|
||||
#print(imgs_ver_flipped, 'imgs_ver_flipped')
|
||||
|
||||
else:
|
||||
imgs_ver_flipped = None
|
||||
|
||||
if img_bin:
|
||||
imgs_bin = cropped_lines_bin[n_start:]
|
||||
imgs_bin = np.array(imgs_bin)
|
||||
imgs_bin = imgs_bin.reshape(imgs_bin.shape[0], image_height, image_width, 3)
|
||||
|
||||
if len(indices_ver)>0:
|
||||
imgs_bin_ver_flipped = imgs_bin[indices_ver, : ,: ,:]
|
||||
imgs_bin_ver_flipped = imgs_bin_ver_flipped[:,::-1,::-1,:]
|
||||
#print(imgs_ver_flipped, 'imgs_ver_flipped')
|
||||
|
||||
else:
|
||||
imgs_bin_ver_flipped = None
|
||||
else:
|
||||
n_start = i*self.b_s
|
||||
n_end = (i+1)*self.b_s
|
||||
imgs = cropped_lines[n_start:n_end]
|
||||
imgs = np.array(imgs).reshape(self.b_s, image_height, image_width, 3)
|
||||
|
||||
ver_imgs = np.array( cropped_lines_ver_index[n_start:n_end] )
|
||||
indices_ver = np.where(ver_imgs == 1)[0]
|
||||
#print(indices_ver, 'indices_ver')
|
||||
|
||||
if len(indices_ver)>0:
|
||||
imgs_ver_flipped = imgs[indices_ver, : ,: ,:]
|
||||
imgs_ver_flipped = imgs_ver_flipped[:,::-1,::-1,:]
|
||||
#print(imgs_ver_flipped, 'imgs_ver_flipped')
|
||||
else:
|
||||
imgs_ver_flipped = None
|
||||
|
||||
|
||||
if img_bin:
|
||||
imgs_bin = cropped_lines_bin[n_start:n_end]
|
||||
imgs_bin = np.array(imgs_bin).reshape(self.b_s, image_height, image_width, 3)
|
||||
|
||||
|
||||
if len(indices_ver)>0:
|
||||
imgs_bin_ver_flipped = imgs_bin[indices_ver, : ,: ,:]
|
||||
imgs_bin_ver_flipped = imgs_bin_ver_flipped[:,::-1,::-1,:]
|
||||
#print(imgs_ver_flipped, 'imgs_ver_flipped')
|
||||
else:
|
||||
imgs_bin_ver_flipped = None
|
||||
|
||||
|
||||
self.logger.debug("processing next %d lines", len(imgs))
|
||||
preds = self.model_zoo.get('ocr').predict(imgs, verbose=0)
|
||||
|
||||
if len(indices_ver)>0:
|
||||
preds_flipped = self.model_zoo.get('ocr').predict(imgs_ver_flipped, verbose=0)
|
||||
preds_max_fliped = np.max(preds_flipped, axis=2 )
|
||||
preds_max_args_flipped = np.argmax(preds_flipped, axis=2 )
|
||||
pred_max_not_unk_mask_bool_flipped = preds_max_args_flipped[:,:]!=self.end_character
|
||||
masked_means_flipped = \
|
||||
np.sum(preds_max_fliped * pred_max_not_unk_mask_bool_flipped, axis=1) / \
|
||||
np.sum(pred_max_not_unk_mask_bool_flipped, axis=1)
|
||||
masked_means_flipped[np.isnan(masked_means_flipped)] = 0
|
||||
|
||||
preds_max = np.max(preds, axis=2 )
|
||||
preds_max_args = np.argmax(preds, axis=2 )
|
||||
pred_max_not_unk_mask_bool = preds_max_args[:,:]!=self.end_character
|
||||
|
||||
masked_means = \
|
||||
np.sum(preds_max * pred_max_not_unk_mask_bool, axis=1) / \
|
||||
np.sum(pred_max_not_unk_mask_bool, axis=1)
|
||||
masked_means[np.isnan(masked_means)] = 0
|
||||
|
||||
masked_means_ver = masked_means[indices_ver]
|
||||
#print(masked_means_ver, 'pred_max_not_unk')
|
||||
|
||||
indices_where_flipped_conf_value_is_higher = \
|
||||
np.where(masked_means_flipped > masked_means_ver)[0]
|
||||
|
||||
#print(indices_where_flipped_conf_value_is_higher, 'indices_where_flipped_conf_value_is_higher')
|
||||
if len(indices_where_flipped_conf_value_is_higher)>0:
|
||||
indices_to_be_replaced = indices_ver[indices_where_flipped_conf_value_is_higher]
|
||||
preds[indices_to_be_replaced,:,:] = \
|
||||
preds_flipped[indices_where_flipped_conf_value_is_higher, :, :]
|
||||
|
||||
if img_bin:
|
||||
preds_bin = self.model_zoo.get('ocr').predict(imgs_bin, verbose=0)
|
||||
|
||||
if len(indices_ver)>0:
|
||||
preds_flipped = self.model_zoo.get('ocr').predict(imgs_bin_ver_flipped, verbose=0)
|
||||
preds_max_fliped = np.max(preds_flipped, axis=2 )
|
||||
preds_max_args_flipped = np.argmax(preds_flipped, axis=2 )
|
||||
pred_max_not_unk_mask_bool_flipped = preds_max_args_flipped[:,:]!=self.end_character
|
||||
masked_means_flipped = \
|
||||
np.sum(preds_max_fliped * pred_max_not_unk_mask_bool_flipped, axis=1) / \
|
||||
np.sum(pred_max_not_unk_mask_bool_flipped, axis=1)
|
||||
masked_means_flipped[np.isnan(masked_means_flipped)] = 0
|
||||
|
||||
preds_max = np.max(preds, axis=2 )
|
||||
preds_max_args = np.argmax(preds, axis=2 )
|
||||
pred_max_not_unk_mask_bool = preds_max_args[:,:]!=self.end_character
|
||||
|
||||
masked_means = \
|
||||
np.sum(preds_max * pred_max_not_unk_mask_bool, axis=1) / \
|
||||
np.sum(pred_max_not_unk_mask_bool, axis=1)
|
||||
masked_means[np.isnan(masked_means)] = 0
|
||||
|
||||
masked_means_ver = masked_means[indices_ver]
|
||||
#print(masked_means_ver, 'pred_max_not_unk')
|
||||
|
||||
indices_where_flipped_conf_value_is_higher = \
|
||||
np.where(masked_means_flipped > masked_means_ver)[0]
|
||||
|
||||
#print(indices_where_flipped_conf_value_is_higher, 'indices_where_flipped_conf_value_is_higher')
|
||||
if len(indices_where_flipped_conf_value_is_higher)>0:
|
||||
indices_to_be_replaced = indices_ver[indices_where_flipped_conf_value_is_higher]
|
||||
preds_bin[indices_to_be_replaced,:,:] = \
|
||||
preds_flipped[indices_where_flipped_conf_value_is_higher, :, :]
|
||||
|
||||
preds = (preds + preds_bin) / 2.
|
||||
|
||||
pred_texts = decode_batch_predictions(preds, self.model_zoo.get('num_to_char'))
|
||||
|
||||
preds_max = np.max(preds, axis=2 )
|
||||
preds_max_args = np.argmax(preds, axis=2 )
|
||||
pred_max_not_unk_mask_bool = preds_max_args[:,:]!=self.end_character
|
||||
masked_means = \
|
||||
np.sum(preds_max * pred_max_not_unk_mask_bool, axis=1) / \
|
||||
np.sum(pred_max_not_unk_mask_bool, axis=1)
|
||||
|
||||
for ib in range(imgs.shape[0]):
|
||||
pred_texts_ib = pred_texts[ib].replace("[UNK]", "")
|
||||
if masked_means[ib] >= self.min_conf_value_of_textline_text:
|
||||
extracted_texts.append(pred_texts_ib)
|
||||
extracted_conf_value.append(masked_means[ib])
|
||||
else:
|
||||
extracted_texts.append("")
|
||||
extracted_conf_value.append(0)
|
||||
del cropped_lines
|
||||
del cropped_lines_bin
|
||||
gc.collect()
|
||||
|
||||
extracted_texts_merged = [extracted_texts[ind]
|
||||
if cropped_lines_meging_indexing[ind]==0
|
||||
else extracted_texts[ind]+" "+extracted_texts[ind+1]
|
||||
if cropped_lines_meging_indexing[ind]==1
|
||||
else None
|
||||
for ind in range(len(cropped_lines_meging_indexing))]
|
||||
|
||||
extracted_conf_value_merged = [extracted_conf_value[ind] # type: ignore
|
||||
if cropped_lines_meging_indexing[ind]==0
|
||||
else (extracted_conf_value[ind]+extracted_conf_value[ind+1])/2.
|
||||
if cropped_lines_meging_indexing[ind]==1
|
||||
else None
|
||||
for ind in range(len(cropped_lines_meging_indexing))]
|
||||
|
||||
extracted_conf_value_merged: List[float] = [extracted_conf_value_merged[ind_cfm]
|
||||
for ind_cfm in range(len(extracted_texts_merged))
|
||||
if extracted_texts_merged[ind_cfm] is not None]
|
||||
|
||||
extracted_texts_merged = [ind for ind in extracted_texts_merged if ind is not None]
|
||||
|
||||
return EynollahOcrResult(
|
||||
extracted_texts_merged=extracted_texts_merged,
|
||||
extracted_conf_value_merged=extracted_conf_value_merged,
|
||||
cropped_lines_region_indexer=cropped_lines_region_indexer,
|
||||
total_bb_coordinates=total_bb_coordinates,
|
||||
)
|
||||
|
||||
def write_ocr(
|
||||
self,
|
||||
*,
|
||||
result: EynollahOcrResult,
|
||||
page_tree: ET.ElementTree,
|
||||
out_file_ocr,
|
||||
page_ns,
|
||||
img,
|
||||
out_image_with_text,
|
||||
):
|
||||
cropped_lines_region_indexer = result.cropped_lines_region_indexer
|
||||
total_bb_coordinates = result.total_bb_coordinates
|
||||
extracted_texts_merged = result.extracted_texts_merged
|
||||
extracted_conf_value_merged = result.extracted_conf_value_merged
|
||||
|
||||
unique_cropped_lines_region_indexer = np.unique(cropped_lines_region_indexer)
|
||||
if out_image_with_text:
|
||||
image_text = Image.new("RGB", (img.shape[1], img.shape[0]), "white")
|
||||
draw = ImageDraw.Draw(image_text)
|
||||
font = get_font()
|
||||
|
||||
for indexer_text, bb_ind in enumerate(total_bb_coordinates):
|
||||
x_bb = bb_ind[0]
|
||||
y_bb = bb_ind[1]
|
||||
w_bb = bb_ind[2]
|
||||
h_bb = bb_ind[3]
|
||||
|
||||
font = fit_text_single_line(draw, extracted_texts_merged[indexer_text],
|
||||
font.path, w_bb, int(h_bb*0.4) )
|
||||
|
||||
##draw.rectangle([x_bb, y_bb, x_bb + w_bb, y_bb + h_bb], outline="red", width=2)
|
||||
|
||||
text_bbox = draw.textbbox((0, 0), extracted_texts_merged[indexer_text], font=font)
|
||||
text_width = text_bbox[2] - text_bbox[0]
|
||||
text_height = text_bbox[3] - text_bbox[1]
|
||||
|
||||
text_x = x_bb + (w_bb - text_width) // 2 # Center horizontally
|
||||
text_y = y_bb + (h_bb - text_height) // 2 # Center vertically
|
||||
|
||||
# Draw the text
|
||||
draw.text((text_x, text_y), extracted_texts_merged[indexer_text], fill="black", font=font)
|
||||
image_text.save(out_image_with_text)
|
||||
|
||||
text_by_textregion = []
|
||||
for ind in unique_cropped_lines_region_indexer:
|
||||
ind = np.array(cropped_lines_region_indexer)==ind
|
||||
extracted_texts_merged_un = np.array(extracted_texts_merged)[ind]
|
||||
if len(extracted_texts_merged_un)>1:
|
||||
text_by_textregion_ind = ""
|
||||
next_glue = ""
|
||||
for indt in range(len(extracted_texts_merged_un)):
|
||||
if (extracted_texts_merged_un[indt].endswith('⸗') or
|
||||
extracted_texts_merged_un[indt].endswith('-') or
|
||||
extracted_texts_merged_un[indt].endswith('¬')):
|
||||
text_by_textregion_ind += next_glue + extracted_texts_merged_un[indt][:-1]
|
||||
next_glue = ""
|
||||
else:
|
||||
text_by_textregion_ind += next_glue + extracted_texts_merged_un[indt]
|
||||
next_glue = " "
|
||||
text_by_textregion.append(text_by_textregion_ind)
|
||||
else:
|
||||
text_by_textregion.append(" ".join(extracted_texts_merged_un))
|
||||
|
||||
indexer = 0
|
||||
indexer_textregion = 0
|
||||
for nn in page_tree.getroot().iter(f'{{{page_ns}}}TextRegion'):
|
||||
|
||||
is_textregion_text = False
|
||||
for childtest in nn:
|
||||
if childtest.tag.endswith("TextEquiv"):
|
||||
is_textregion_text = True
|
||||
|
||||
if not is_textregion_text:
|
||||
text_subelement_textregion = ET.SubElement(nn, 'TextEquiv')
|
||||
unicode_textregion = ET.SubElement(text_subelement_textregion, 'Unicode')
|
||||
|
||||
|
||||
has_textline = False
|
||||
for child_textregion in nn:
|
||||
if child_textregion.tag.endswith("TextLine"):
|
||||
|
||||
is_textline_text = False
|
||||
for childtest2 in child_textregion:
|
||||
if childtest2.tag.endswith("TextEquiv"):
|
||||
is_textline_text = True
|
||||
|
||||
|
||||
if not is_textline_text:
|
||||
text_subelement = ET.SubElement(child_textregion, 'TextEquiv')
|
||||
if extracted_conf_value_merged:
|
||||
text_subelement.set('conf', f"{extracted_conf_value_merged[indexer]:.2f}")
|
||||
unicode_textline = ET.SubElement(text_subelement, 'Unicode')
|
||||
unicode_textline.text = extracted_texts_merged[indexer]
|
||||
else:
|
||||
for childtest3 in child_textregion:
|
||||
if childtest3.tag.endswith("TextEquiv"):
|
||||
for child_uc in childtest3:
|
||||
if child_uc.tag.endswith("Unicode"):
|
||||
if extracted_conf_value_merged:
|
||||
childtest3.set('conf', f"{extracted_conf_value_merged[indexer]:.2f}")
|
||||
child_uc.text = extracted_texts_merged[indexer]
|
||||
|
||||
indexer = indexer + 1
|
||||
has_textline = True
|
||||
if has_textline:
|
||||
if is_textregion_text:
|
||||
for child4 in nn:
|
||||
if child4.tag.endswith("TextEquiv"):
|
||||
for childtr_uc in child4:
|
||||
if childtr_uc.tag.endswith("Unicode"):
|
||||
childtr_uc.text = text_by_textregion[indexer_textregion]
|
||||
else:
|
||||
unicode_textregion.text = text_by_textregion[indexer_textregion]
|
||||
indexer_textregion = indexer_textregion + 1
|
||||
|
||||
ET.register_namespace("",page_ns)
|
||||
page_tree.write(out_file_ocr, xml_declaration=True, method='xml', encoding="utf-8", default_namespace=None)
|
||||
|
||||
def run(
|
||||
self,
|
||||
*,
|
||||
overwrite: bool = False,
|
||||
dir_in: Optional[str] = None,
|
||||
dir_in_bin: Optional[str] = None,
|
||||
image_filename: Optional[str] = None,
|
||||
dir_xmls: str,
|
||||
dir_out_image_text: Optional[str] = None,
|
||||
dir_out: str,
|
||||
):
|
||||
"""
|
||||
Run OCR.
|
||||
|
||||
Args:
|
||||
|
||||
dir_in_bin (str): Prediction with RGB and binarized images for selected pages, should not be the default
|
||||
"""
|
||||
if dir_in:
|
||||
ls_imgs = [os.path.join(dir_in, image_filename)
|
||||
for image_filename in filter(is_image_filename,
|
||||
os.listdir(dir_in))]
|
||||
else:
|
||||
assert image_filename
|
||||
ls_imgs = [image_filename]
|
||||
|
||||
for img_filename in ls_imgs:
|
||||
file_stem = Path(img_filename).stem
|
||||
page_file_in = os.path.join(dir_xmls, file_stem+'.xml')
|
||||
out_file_ocr = os.path.join(dir_out, file_stem+'.xml')
|
||||
|
||||
if os.path.exists(out_file_ocr):
|
||||
if overwrite:
|
||||
self.logger.warning("will overwrite existing output file '%s'", out_file_ocr)
|
||||
else:
|
||||
self.logger.warning("will skip input for existing output file '%s'", out_file_ocr)
|
||||
return
|
||||
|
||||
img = cv2.imread(img_filename)
|
||||
|
||||
page_tree = ET.parse(page_file_in, parser = ET.XMLParser(encoding="utf-8"))
|
||||
page_ns = etree_namespace_for_element_tag(page_tree.getroot().tag)
|
||||
|
||||
out_image_with_text = None
|
||||
if dir_out_image_text:
|
||||
out_image_with_text = os.path.join(dir_out_image_text, file_stem + '.png')
|
||||
|
||||
img_bin = None
|
||||
if dir_in_bin:
|
||||
img_bin = cv2.imread(os.path.join(dir_in_bin, file_stem+'.png'))
|
||||
|
||||
|
||||
if self.tr_ocr:
|
||||
result = self.run_trocr(
|
||||
img=img,
|
||||
page_tree=page_tree,
|
||||
page_ns=page_ns,
|
||||
|
||||
tr_ocr_input_height_and_width = 384
|
||||
)
|
||||
else:
|
||||
result = self.run_cnn(
|
||||
img=img,
|
||||
page_tree=page_tree,
|
||||
page_ns=page_ns,
|
||||
|
||||
img_bin=img_bin,
|
||||
image_width=512,
|
||||
image_height=32,
|
||||
)
|
||||
|
||||
self.write_ocr(
|
||||
result=result,
|
||||
img=img,
|
||||
page_tree=page_tree,
|
||||
page_ns=page_ns,
|
||||
out_file_ocr=out_file_ocr,
|
||||
out_image_with_text=out_image_with_text,
|
||||
)
|
||||
|
|
@ -1,94 +0,0 @@
|
|||
"""
|
||||
Image enhancer. The output can be written as same scale of input or in new predicted scale.
|
||||
"""
|
||||
|
||||
import logging
|
||||
import os
|
||||
from typing import Optional
|
||||
from pathlib import Path
|
||||
|
||||
import cv2
|
||||
|
||||
from .eynollah import Eynollah
|
||||
from .model_zoo import EynollahModelZoo
|
||||
from .utils.resize import resize_image
|
||||
from .utils import is_image_filename
|
||||
|
||||
|
||||
class Enhancer(Eynollah):
|
||||
def __init__(
|
||||
self,
|
||||
*,
|
||||
model_zoo: EynollahModelZoo,
|
||||
num_col_upper: int = 0,
|
||||
num_col_lower: int = 0,
|
||||
save_org_scale: bool = False,
|
||||
device: str = '',
|
||||
):
|
||||
self.save_org_scale = save_org_scale
|
||||
self.num_col_upper = int(num_col_upper)
|
||||
self.num_col_lower = int(num_col_lower)
|
||||
self.input_binary = False
|
||||
self.ignore_page_extraction = False
|
||||
|
||||
self.logger = logging.getLogger('eynollah.enhance')
|
||||
self.model_zoo = model_zoo
|
||||
self.setup_models(device=device)
|
||||
|
||||
def setup_models(self, device=''):
|
||||
loadable = ['enhancement', 'col_classifier', 'page']
|
||||
self.model_zoo.load_models(*loadable, device=device)
|
||||
for model in loadable:
|
||||
self.logger.debug("model %s has input shape %s", model,
|
||||
self.model_zoo.get(model).input_shape)
|
||||
|
||||
def run_single(self,
|
||||
img_filename: str,
|
||||
img_pil=None,
|
||||
dir_out: Optional[str] = None,
|
||||
overwrite: bool = False,
|
||||
) -> None:
|
||||
|
||||
image = self.cache_images(image_filename=img_filename, image_pil=img_pil)
|
||||
output_filename = os.path.join(dir_out or "", image['name'] + '.png')
|
||||
|
||||
if os.path.exists(output_filename):
|
||||
if overwrite:
|
||||
self.logger.warning("will overwrite existing output file '%s'", output_filename)
|
||||
else:
|
||||
self.logger.warning("will skip input for existing output file '%s'", output_filename)
|
||||
return
|
||||
|
||||
self.resize_image_with_column_classifier(image)
|
||||
img_org = image['img']
|
||||
img_res = image['img_res']
|
||||
if self.save_org_scale:
|
||||
img_res = resize_image(img_res, img_org.shape[0], img_org.shape[1])
|
||||
|
||||
cv2.imwrite(output_filename, img_res)
|
||||
self.logger.info("output filename: '%s'", output_filename)
|
||||
|
||||
def run(self,
|
||||
overwrite: bool = False,
|
||||
image_filename: Optional[str] = None,
|
||||
dir_in: Optional[str] = None,
|
||||
dir_out: Optional[str] = None,
|
||||
):
|
||||
"""
|
||||
Enlarge and enhance the scanned images
|
||||
"""
|
||||
if dir_in:
|
||||
ls_imgs = [os.path.join(dir_in, image_filename)
|
||||
for image_filename in filter(is_image_filename,
|
||||
os.listdir(dir_in))]
|
||||
elif image_filename:
|
||||
ls_imgs = [image_filename]
|
||||
else:
|
||||
raise ValueError("run requires either a single image filename or a directory")
|
||||
|
||||
for img_filename in ls_imgs:
|
||||
self.logger.info(img_filename)
|
||||
|
||||
self.run_single(img_filename,
|
||||
dir_out=dir_out,
|
||||
overwrite=overwrite)
|
||||
|
|
@ -1,805 +0,0 @@
|
|||
"""
|
||||
Machine learning based reading order detection
|
||||
"""
|
||||
|
||||
# pyright: reportCallIssue=false
|
||||
# pyright: reportUnboundVariable=false
|
||||
# pyright: reportArgumentType=false
|
||||
|
||||
import logging
|
||||
import os
|
||||
import time
|
||||
from typing import Optional
|
||||
from pathlib import Path
|
||||
import xml.etree.ElementTree as ET
|
||||
|
||||
import cv2
|
||||
import numpy as np
|
||||
import statistics
|
||||
|
||||
os.environ['TF_USE_LEGACY_KERAS'] = '1' # avoid Keras 3 after TF 2.15
|
||||
import tensorflow as tf
|
||||
|
||||
from .model_zoo import EynollahModelZoo
|
||||
from .utils.resize import resize_image
|
||||
from .utils.contour import (
|
||||
find_new_features_of_contours,
|
||||
return_contours_of_image,
|
||||
return_parent_contours,
|
||||
)
|
||||
from .utils import is_xml_filename
|
||||
|
||||
DPI_THRESHOLD = 298
|
||||
KERNEL = np.ones((5, 5), np.uint8)
|
||||
|
||||
|
||||
class machine_based_reading_order_on_layout:
|
||||
def __init__(
|
||||
self,
|
||||
*,
|
||||
model_zoo: EynollahModelZoo,
|
||||
logger : Optional[logging.Logger] = None,
|
||||
):
|
||||
self.logger = logger or logging.getLogger('eynollah.mbreorder')
|
||||
self.model_zoo = model_zoo
|
||||
|
||||
try:
|
||||
for device in tf.config.list_physical_devices('GPU'):
|
||||
tf.config.experimental.set_memory_growth(device, True)
|
||||
except:
|
||||
self.logger.warning("no GPU device available")
|
||||
|
||||
self.model_zoo.load_models('reading_order')
|
||||
|
||||
def read_xml(self, xml_file):
|
||||
tree1 = ET.parse(xml_file, parser = ET.XMLParser(encoding='utf-8'))
|
||||
root1=tree1.getroot()
|
||||
alltags=[elem.tag for elem in root1.iter()]
|
||||
link=alltags[0].split('}')[0]+'}'
|
||||
|
||||
index_tot_regions = []
|
||||
tot_region_ref = []
|
||||
|
||||
y_len, x_len = 0, 0
|
||||
for jj in root1.iter(link+'Page'):
|
||||
y_len=int(jj.attrib['imageHeight'])
|
||||
x_len=int(jj.attrib['imageWidth'])
|
||||
|
||||
for jj in root1.iter(link+'RegionRefIndexed'):
|
||||
index_tot_regions.append(jj.attrib['index'])
|
||||
tot_region_ref.append(jj.attrib['regionRef'])
|
||||
|
||||
if (link+'PrintSpace' in alltags) or (link+'Border' in alltags):
|
||||
co_printspace = []
|
||||
if link+'PrintSpace' in alltags:
|
||||
region_tags_printspace = np.unique([x for x in alltags if x.endswith('PrintSpace')])
|
||||
else:
|
||||
region_tags_printspace = np.unique([x for x in alltags if x.endswith('Border')])
|
||||
|
||||
for tag in region_tags_printspace:
|
||||
if link+'PrintSpace' in alltags:
|
||||
tag_endings_printspace = ['}PrintSpace','}printspace']
|
||||
else:
|
||||
tag_endings_printspace = ['}Border','}border']
|
||||
|
||||
if tag.endswith(tag_endings_printspace[0]) or tag.endswith(tag_endings_printspace[1]):
|
||||
for nn in root1.iter(tag):
|
||||
c_t_in = []
|
||||
sumi = 0
|
||||
for vv in nn.iter():
|
||||
# check the format of coords
|
||||
if vv.tag == link + 'Coords':
|
||||
coords = bool(vv.attrib)
|
||||
if coords:
|
||||
p_h = vv.attrib['points'].split(' ')
|
||||
c_t_in.append(
|
||||
np.array([[int(x.split(',')[0]), int(x.split(',')[1])] for x in p_h]))
|
||||
break
|
||||
else:
|
||||
pass
|
||||
|
||||
if vv.tag == link + 'Point':
|
||||
c_t_in.append([int(float(vv.attrib['x'])), int(float(vv.attrib['y']))])
|
||||
sumi += 1
|
||||
elif vv.tag != link + 'Point' and sumi >= 1:
|
||||
break
|
||||
co_printspace.append(np.array(c_t_in))
|
||||
img_printspace = np.zeros( (y_len,x_len,3) )
|
||||
img_printspace=cv2.fillPoly(img_printspace, pts =co_printspace, color=(1,1,1))
|
||||
img_printspace = img_printspace.astype(np.uint8)
|
||||
|
||||
imgray = cv2.cvtColor(img_printspace, cv2.COLOR_BGR2GRAY)
|
||||
_, thresh = cv2.threshold(imgray, 0, 255, 0)
|
||||
contours, _ = cv2.findContours(thresh, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
|
||||
cnt_size = np.array([cv2.contourArea(contours[j]) for j in range(len(contours))])
|
||||
cnt = contours[np.argmax(cnt_size)]
|
||||
x, y, w, h = cv2.boundingRect(cnt)
|
||||
|
||||
bb_coord_printspace = [x, y, w, h]
|
||||
|
||||
else:
|
||||
bb_coord_printspace = None
|
||||
|
||||
|
||||
region_tags=np.unique([x for x in alltags if x.endswith('Region')])
|
||||
co_text_paragraph=[]
|
||||
co_text_drop=[]
|
||||
co_text_heading=[]
|
||||
co_text_header=[]
|
||||
co_text_marginalia=[]
|
||||
co_text_catch=[]
|
||||
co_text_page_number=[]
|
||||
co_text_signature_mark=[]
|
||||
co_sep=[]
|
||||
co_img=[]
|
||||
co_table=[]
|
||||
co_graphic=[]
|
||||
co_graphic_text_annotation=[]
|
||||
co_graphic_decoration=[]
|
||||
co_noise=[]
|
||||
|
||||
co_text_paragraph_text=[]
|
||||
co_text_drop_text=[]
|
||||
co_text_heading_text=[]
|
||||
co_text_header_text=[]
|
||||
co_text_marginalia_text=[]
|
||||
co_text_catch_text=[]
|
||||
co_text_page_number_text=[]
|
||||
co_text_signature_mark_text=[]
|
||||
co_sep_text=[]
|
||||
co_img_text=[]
|
||||
co_table_text=[]
|
||||
co_graphic_text=[]
|
||||
co_graphic_text_annotation_text=[]
|
||||
co_graphic_decoration_text=[]
|
||||
co_noise_text=[]
|
||||
|
||||
id_paragraph = []
|
||||
id_header = []
|
||||
id_heading = []
|
||||
id_marginalia = []
|
||||
|
||||
for tag in region_tags:
|
||||
if tag.endswith('}TextRegion') or tag.endswith('}Textregion'):
|
||||
for nn in root1.iter(tag):
|
||||
for child2 in nn:
|
||||
tag2 = child2.tag
|
||||
if tag2.endswith('}TextEquiv') or tag2.endswith('}TextEquiv'):
|
||||
for childtext2 in child2:
|
||||
if childtext2.tag.endswith('}Unicode') or childtext2.tag.endswith('}Unicode'):
|
||||
if "type" in nn.attrib and nn.attrib['type']=='drop-capital':
|
||||
co_text_drop_text.append(childtext2.text)
|
||||
elif "type" in nn.attrib and nn.attrib['type']=='heading':
|
||||
co_text_heading_text.append(childtext2.text)
|
||||
elif "type" in nn.attrib and nn.attrib['type']=='signature-mark':
|
||||
co_text_signature_mark_text.append(childtext2.text)
|
||||
elif "type" in nn.attrib and nn.attrib['type']=='header':
|
||||
co_text_header_text.append(childtext2.text)
|
||||
###elif "type" in nn.attrib and nn.attrib['type']=='catch-word':
|
||||
###co_text_catch_text.append(childtext2.text)
|
||||
###elif "type" in nn.attrib and nn.attrib['type']=='page-number':
|
||||
###co_text_page_number_text.append(childtext2.text)
|
||||
elif "type" in nn.attrib and nn.attrib['type']=='marginalia':
|
||||
co_text_marginalia_text.append(childtext2.text)
|
||||
else:
|
||||
co_text_paragraph_text.append(childtext2.text)
|
||||
c_t_in_drop=[]
|
||||
c_t_in_paragraph=[]
|
||||
c_t_in_heading=[]
|
||||
c_t_in_header=[]
|
||||
c_t_in_page_number=[]
|
||||
c_t_in_signature_mark=[]
|
||||
c_t_in_catch=[]
|
||||
c_t_in_marginalia=[]
|
||||
|
||||
|
||||
sumi=0
|
||||
for vv in nn.iter():
|
||||
# check the format of coords
|
||||
if vv.tag==link+'Coords':
|
||||
|
||||
coords=bool(vv.attrib)
|
||||
if coords:
|
||||
#print('birda1')
|
||||
p_h=vv.attrib['points'].split(' ')
|
||||
|
||||
|
||||
|
||||
if "type" in nn.attrib and nn.attrib['type']=='drop-capital':
|
||||
|
||||
c_t_in_drop.append( np.array( [ [ int(x.split(',')[0]) , int(x.split(',')[1]) ] for x in p_h] ) )
|
||||
|
||||
elif "type" in nn.attrib and nn.attrib['type']=='heading':
|
||||
##id_heading.append(nn.attrib['id'])
|
||||
c_t_in_heading.append( np.array( [ [ int(x.split(',')[0]) , int(x.split(',')[1]) ] for x in p_h] ) )
|
||||
|
||||
|
||||
elif "type" in nn.attrib and nn.attrib['type']=='signature-mark':
|
||||
|
||||
c_t_in_signature_mark.append( np.array( [ [ int(x.split(',')[0]) , int(x.split(',')[1]) ] for x in p_h] ) )
|
||||
#print(c_t_in_paragraph)
|
||||
elif "type" in nn.attrib and nn.attrib['type']=='header':
|
||||
#id_header.append(nn.attrib['id'])
|
||||
c_t_in_header.append( np.array( [ [ int(x.split(',')[0]) , int(x.split(',')[1]) ] for x in p_h] ) )
|
||||
|
||||
|
||||
###elif "type" in nn.attrib and nn.attrib['type']=='catch-word':
|
||||
###c_t_in_catch.append( np.array( [ [ int(x.split(',')[0]) , int(x.split(',')[1]) ] for x in p_h] ) )
|
||||
|
||||
|
||||
###elif "type" in nn.attrib and nn.attrib['type']=='page-number':
|
||||
|
||||
###c_t_in_page_number.append( np.array( [ [ int(x.split(',')[0]) , int(x.split(',')[1]) ] for x in p_h] ) )
|
||||
|
||||
elif "type" in nn.attrib and nn.attrib['type']=='marginalia':
|
||||
#id_marginalia.append(nn.attrib['id'])
|
||||
|
||||
c_t_in_marginalia.append( np.array( [ [ int(x.split(',')[0]) , int(x.split(',')[1]) ] for x in p_h] ) )
|
||||
else:
|
||||
#id_paragraph.append(nn.attrib['id'])
|
||||
|
||||
c_t_in_paragraph.append( np.array( [ [ int(x.split(',')[0]) , int(x.split(',')[1]) ] for x in p_h] ) )
|
||||
|
||||
break
|
||||
else:
|
||||
pass
|
||||
|
||||
|
||||
if vv.tag==link+'Point':
|
||||
if "type" in nn.attrib and nn.attrib['type']=='drop-capital':
|
||||
|
||||
c_t_in_drop.append([ int(float(vv.attrib['x'])) , int(float(vv.attrib['y'])) ])
|
||||
sumi+=1
|
||||
|
||||
elif "type" in nn.attrib and nn.attrib['type']=='heading':
|
||||
#id_heading.append(nn.attrib['id'])
|
||||
c_t_in_heading.append([ int(float(vv.attrib['x'])) , int(float(vv.attrib['y'])) ])
|
||||
sumi+=1
|
||||
|
||||
|
||||
elif "type" in nn.attrib and nn.attrib['type']=='signature-mark':
|
||||
|
||||
c_t_in_signature_mark.append([ int(float(vv.attrib['x'])) , int(float(vv.attrib['y'])) ])
|
||||
sumi+=1
|
||||
elif "type" in nn.attrib and nn.attrib['type']=='header':
|
||||
#id_header.append(nn.attrib['id'])
|
||||
c_t_in_header.append([ int(float(vv.attrib['x'])) , int(float(vv.attrib['y'])) ])
|
||||
sumi+=1
|
||||
|
||||
|
||||
###elif "type" in nn.attrib and nn.attrib['type']=='catch-word':
|
||||
###c_t_in_catch.append([ int(float(vv.attrib['x'])) , int(float(vv.attrib['y'])) ])
|
||||
###sumi+=1
|
||||
|
||||
###elif "type" in nn.attrib and nn.attrib['type']=='page-number':
|
||||
|
||||
###c_t_in_page_number.append([ int(float(vv.attrib['x'])) , int(float(vv.attrib['y'])) ])
|
||||
###sumi+=1
|
||||
|
||||
elif "type" in nn.attrib and nn.attrib['type']=='marginalia':
|
||||
#id_marginalia.append(nn.attrib['id'])
|
||||
|
||||
c_t_in_marginalia.append([ int(float(vv.attrib['x'])) , int(float(vv.attrib['y'])) ])
|
||||
sumi+=1
|
||||
|
||||
else:
|
||||
#id_paragraph.append(nn.attrib['id'])
|
||||
c_t_in_paragraph.append([ int(float(vv.attrib['x'])) , int(float(vv.attrib['y'])) ])
|
||||
sumi+=1
|
||||
|
||||
elif vv.tag!=link+'Point' and sumi>=1:
|
||||
break
|
||||
|
||||
if len(c_t_in_drop)>0:
|
||||
co_text_drop.append(np.array(c_t_in_drop))
|
||||
if len(c_t_in_paragraph)>0:
|
||||
co_text_paragraph.append(np.array(c_t_in_paragraph))
|
||||
id_paragraph.append(nn.attrib['id'])
|
||||
if len(c_t_in_heading)>0:
|
||||
co_text_heading.append(np.array(c_t_in_heading))
|
||||
id_heading.append(nn.attrib['id'])
|
||||
|
||||
if len(c_t_in_header)>0:
|
||||
co_text_header.append(np.array(c_t_in_header))
|
||||
id_header.append(nn.attrib['id'])
|
||||
if len(c_t_in_page_number)>0:
|
||||
co_text_page_number.append(np.array(c_t_in_page_number))
|
||||
if len(c_t_in_catch)>0:
|
||||
co_text_catch.append(np.array(c_t_in_catch))
|
||||
|
||||
if len(c_t_in_signature_mark)>0:
|
||||
co_text_signature_mark.append(np.array(c_t_in_signature_mark))
|
||||
|
||||
if len(c_t_in_marginalia)>0:
|
||||
co_text_marginalia.append(np.array(c_t_in_marginalia))
|
||||
id_marginalia.append(nn.attrib['id'])
|
||||
|
||||
|
||||
elif tag.endswith('}GraphicRegion') or tag.endswith('}graphicregion'):
|
||||
for nn in root1.iter(tag):
|
||||
c_t_in=[]
|
||||
c_t_in_text_annotation=[]
|
||||
c_t_in_decoration=[]
|
||||
sumi=0
|
||||
for vv in nn.iter():
|
||||
# check the format of coords
|
||||
if vv.tag==link+'Coords':
|
||||
coords=bool(vv.attrib)
|
||||
if coords:
|
||||
p_h=vv.attrib['points'].split(' ')
|
||||
|
||||
if "type" in nn.attrib and nn.attrib['type']=='handwritten-annotation':
|
||||
c_t_in_text_annotation.append( np.array( [ [ int(x.split(',')[0]) , int(x.split(',')[1]) ] for x in p_h] ) )
|
||||
|
||||
elif "type" in nn.attrib and nn.attrib['type']=='decoration':
|
||||
c_t_in_decoration.append( np.array( [ [ int(x.split(',')[0]) , int(x.split(',')[1]) ] for x in p_h] ) )
|
||||
|
||||
else:
|
||||
c_t_in.append( np.array( [ [ int(x.split(',')[0]) , int(x.split(',')[1]) ] for x in p_h] ) )
|
||||
|
||||
|
||||
break
|
||||
else:
|
||||
pass
|
||||
|
||||
|
||||
if vv.tag==link+'Point':
|
||||
if "type" in nn.attrib and nn.attrib['type']=='handwritten-annotation':
|
||||
c_t_in_text_annotation.append([ int(float(vv.attrib['x'])) , int(float(vv.attrib['y'])) ])
|
||||
sumi+=1
|
||||
|
||||
elif "type" in nn.attrib and nn.attrib['type']=='decoration':
|
||||
c_t_in_decoration.append([ int(float(vv.attrib['x'])) , int(float(vv.attrib['y'])) ])
|
||||
sumi+=1
|
||||
|
||||
else:
|
||||
c_t_in.append([ int(float(vv.attrib['x'])) , int(float(vv.attrib['y'])) ])
|
||||
sumi+=1
|
||||
|
||||
if len(c_t_in_text_annotation)>0:
|
||||
co_graphic_text_annotation.append(np.array(c_t_in_text_annotation))
|
||||
if len(c_t_in_decoration)>0:
|
||||
co_graphic_decoration.append(np.array(c_t_in_decoration))
|
||||
if len(c_t_in)>0:
|
||||
co_graphic.append(np.array(c_t_in))
|
||||
|
||||
|
||||
|
||||
elif tag.endswith('}ImageRegion') or tag.endswith('}imageregion'):
|
||||
for nn in root1.iter(tag):
|
||||
c_t_in=[]
|
||||
sumi=0
|
||||
for vv in nn.iter():
|
||||
# check the format of coords
|
||||
if vv.tag==link+'Coords':
|
||||
coords=bool(vv.attrib)
|
||||
if coords:
|
||||
p_h=vv.attrib['points'].split(' ')
|
||||
c_t_in.append( np.array( [ [ int(x.split(',')[0]) , int(x.split(',')[1]) ] for x in p_h] ) )
|
||||
break
|
||||
else:
|
||||
pass
|
||||
|
||||
|
||||
if vv.tag==link+'Point':
|
||||
c_t_in.append([ int(float(vv.attrib['x'])) , int(float(vv.attrib['y'])) ])
|
||||
sumi+=1
|
||||
elif vv.tag!=link+'Point' and sumi>=1:
|
||||
break
|
||||
co_img.append(np.array(c_t_in))
|
||||
co_img_text.append(' ')
|
||||
|
||||
|
||||
elif tag.endswith('}SeparatorRegion') or tag.endswith('}separatorregion'):
|
||||
for nn in root1.iter(tag):
|
||||
c_t_in=[]
|
||||
sumi=0
|
||||
for vv in nn.iter():
|
||||
# check the format of coords
|
||||
if vv.tag==link+'Coords':
|
||||
coords=bool(vv.attrib)
|
||||
if coords:
|
||||
p_h=vv.attrib['points'].split(' ')
|
||||
c_t_in.append( np.array( [ [ int(x.split(',')[0]) , int(x.split(',')[1]) ] for x in p_h] ) )
|
||||
break
|
||||
else:
|
||||
pass
|
||||
|
||||
|
||||
if vv.tag==link+'Point':
|
||||
c_t_in.append([ int(float(vv.attrib['x'])) , int(float(vv.attrib['y'])) ])
|
||||
sumi+=1
|
||||
elif vv.tag!=link+'Point' and sumi>=1:
|
||||
break
|
||||
co_sep.append(np.array(c_t_in))
|
||||
|
||||
|
||||
|
||||
elif tag.endswith('}TableRegion') or tag.endswith('}tableregion'):
|
||||
for nn in root1.iter(tag):
|
||||
c_t_in=[]
|
||||
sumi=0
|
||||
for vv in nn.iter():
|
||||
# check the format of coords
|
||||
if vv.tag==link+'Coords':
|
||||
coords=bool(vv.attrib)
|
||||
if coords:
|
||||
p_h=vv.attrib['points'].split(' ')
|
||||
c_t_in.append( np.array( [ [ int(x.split(',')[0]) , int(x.split(',')[1]) ] for x in p_h] ) )
|
||||
break
|
||||
else:
|
||||
pass
|
||||
|
||||
|
||||
if vv.tag==link+'Point':
|
||||
c_t_in.append([ int(float(vv.attrib['x'])) , int(float(vv.attrib['y'])) ])
|
||||
sumi+=1
|
||||
|
||||
elif vv.tag!=link+'Point' and sumi>=1:
|
||||
break
|
||||
co_table.append(np.array(c_t_in))
|
||||
co_table_text.append(' ')
|
||||
|
||||
elif tag.endswith('}NoiseRegion') or tag.endswith('}noiseregion'):
|
||||
for nn in root1.iter(tag):
|
||||
c_t_in=[]
|
||||
sumi=0
|
||||
for vv in nn.iter():
|
||||
# check the format of coords
|
||||
if vv.tag==link+'Coords':
|
||||
coords=bool(vv.attrib)
|
||||
if coords:
|
||||
p_h=vv.attrib['points'].split(' ')
|
||||
c_t_in.append( np.array( [ [ int(x.split(',')[0]) , int(x.split(',')[1]) ] for x in p_h] ) )
|
||||
break
|
||||
else:
|
||||
pass
|
||||
|
||||
|
||||
if vv.tag==link+'Point':
|
||||
c_t_in.append([ int(float(vv.attrib['x'])) , int(float(vv.attrib['y'])) ])
|
||||
sumi+=1
|
||||
|
||||
elif vv.tag!=link+'Point' and sumi>=1:
|
||||
break
|
||||
co_noise.append(np.array(c_t_in))
|
||||
co_noise_text.append(' ')
|
||||
|
||||
img = np.zeros( (y_len,x_len,3) )
|
||||
img_poly=cv2.fillPoly(img, pts =co_text_paragraph, color=(1,1,1))
|
||||
|
||||
img_poly=cv2.fillPoly(img, pts =co_text_heading, color=(2,2,2))
|
||||
img_poly=cv2.fillPoly(img, pts =co_text_header, color=(2,2,2))
|
||||
img_poly=cv2.fillPoly(img, pts =co_text_marginalia, color=(3,3,3))
|
||||
img_poly=cv2.fillPoly(img, pts =co_img, color=(4,4,4))
|
||||
img_poly=cv2.fillPoly(img, pts =co_sep, color=(5,5,5))
|
||||
|
||||
return tree1, root1, bb_coord_printspace, id_paragraph, id_header+id_heading, co_text_paragraph, co_text_header+co_text_heading,\
|
||||
tot_region_ref,x_len, y_len,index_tot_regions, img_poly
|
||||
|
||||
def return_indexes_of_contours_loctaed_inside_another_list_of_contours(self, contours, contours_loc, cx_main_loc, cy_main_loc, indexes_loc):
|
||||
indexes_of_located_cont = []
|
||||
center_x_coordinates_of_located = []
|
||||
center_y_coordinates_of_located = []
|
||||
#M_main_tot = [cv2.moments(contours_loc[j])
|
||||
#for j in range(len(contours_loc))]
|
||||
#cx_main_loc = [(M_main_tot[j]["m10"] / (M_main_tot[j]["m00"] + 1e-32)) for j in range(len(M_main_tot))]
|
||||
#cy_main_loc = [(M_main_tot[j]["m01"] / (M_main_tot[j]["m00"] + 1e-32)) for j in range(len(M_main_tot))]
|
||||
|
||||
for ij in range(len(contours)):
|
||||
results = [cv2.pointPolygonTest(contours[ij], (cx_main_loc[ind], cy_main_loc[ind]), False)
|
||||
for ind in range(len(cy_main_loc)) ]
|
||||
results = np.array(results)
|
||||
indexes_in = np.where((results == 0) | (results == 1))
|
||||
indexes = indexes_loc[indexes_in]# [(results == 0) | (results == 1)]#np.where((results == 0) | (results == 1))
|
||||
|
||||
indexes_of_located_cont.append(indexes)
|
||||
center_x_coordinates_of_located.append(np.array(cx_main_loc)[indexes_in] )
|
||||
center_y_coordinates_of_located.append(np.array(cy_main_loc)[indexes_in] )
|
||||
|
||||
return indexes_of_located_cont, center_x_coordinates_of_located, center_y_coordinates_of_located
|
||||
|
||||
def do_order_of_regions_with_model(self, contours_only_text_parent, contours_only_text_parent_h, text_regions_p):
|
||||
height1 =672#448
|
||||
width1 = 448#224
|
||||
|
||||
height2 =672#448
|
||||
width2= 448#224
|
||||
|
||||
height3 =672#448
|
||||
width3 = 448#224
|
||||
|
||||
inference_bs = 3
|
||||
|
||||
ver_kernel = np.ones((5, 1), dtype=np.uint8)
|
||||
hor_kernel = np.ones((1, 5), dtype=np.uint8)
|
||||
|
||||
|
||||
min_cont_size_to_be_dilated = 10
|
||||
if len(contours_only_text_parent)>min_cont_size_to_be_dilated:
|
||||
cx_conts, cy_conts, x_min_conts, x_max_conts, y_min_conts, y_max_conts, _ = find_new_features_of_contours(contours_only_text_parent)
|
||||
args_cont_located = np.array(range(len(contours_only_text_parent)))
|
||||
|
||||
diff_y_conts = np.abs(y_max_conts[:]-y_min_conts)
|
||||
diff_x_conts = np.abs(x_max_conts[:]-x_min_conts)
|
||||
|
||||
mean_x = statistics.mean(diff_x_conts)
|
||||
median_x = statistics.median(diff_x_conts)
|
||||
|
||||
|
||||
diff_x_ratio= diff_x_conts/mean_x
|
||||
|
||||
args_cont_located_excluded = args_cont_located[diff_x_ratio>=1.3]
|
||||
args_cont_located_included = args_cont_located[diff_x_ratio<1.3]
|
||||
|
||||
contours_only_text_parent_excluded = [contours_only_text_parent[ind] for ind in range(len(contours_only_text_parent)) if diff_x_ratio[ind]>=1.3]#contours_only_text_parent[diff_x_ratio>=1.3]
|
||||
contours_only_text_parent_included = [contours_only_text_parent[ind] for ind in range(len(contours_only_text_parent)) if diff_x_ratio[ind]<1.3]#contours_only_text_parent[diff_x_ratio<1.3]
|
||||
|
||||
|
||||
cx_conts_excluded = [cx_conts[ind] for ind in range(len(cx_conts)) if diff_x_ratio[ind]>=1.3]#cx_conts[diff_x_ratio>=1.3]
|
||||
cx_conts_included = [cx_conts[ind] for ind in range(len(cx_conts)) if diff_x_ratio[ind]<1.3]#cx_conts[diff_x_ratio<1.3]
|
||||
|
||||
cy_conts_excluded = [cy_conts[ind] for ind in range(len(cy_conts)) if diff_x_ratio[ind]>=1.3]#cy_conts[diff_x_ratio>=1.3]
|
||||
cy_conts_included = [cy_conts[ind] for ind in range(len(cy_conts)) if diff_x_ratio[ind]<1.3]#cy_conts[diff_x_ratio<1.3]
|
||||
|
||||
#print(diff_x_ratio, 'ratio')
|
||||
text_regions_p = text_regions_p.astype('uint8')
|
||||
|
||||
if len(contours_only_text_parent_excluded)>0:
|
||||
textregion_par = np.zeros((text_regions_p.shape[0], text_regions_p.shape[1])).astype('uint8')
|
||||
textregion_par = cv2.fillPoly(textregion_par, pts=contours_only_text_parent_included, color=(1,1))
|
||||
else:
|
||||
textregion_par = (text_regions_p[:,:]==1)*1
|
||||
textregion_par = textregion_par.astype('uint8')
|
||||
|
||||
text_regions_p_textregions_dilated = cv2.erode(textregion_par , hor_kernel, iterations=2)
|
||||
text_regions_p_textregions_dilated = cv2.dilate(text_regions_p_textregions_dilated , ver_kernel, iterations=4)
|
||||
text_regions_p_textregions_dilated = cv2.erode(text_regions_p_textregions_dilated , hor_kernel, iterations=1)
|
||||
text_regions_p_textregions_dilated = cv2.dilate(text_regions_p_textregions_dilated , ver_kernel, iterations=5)
|
||||
text_regions_p_textregions_dilated[text_regions_p[:,:]>1] = 0
|
||||
|
||||
|
||||
contours_only_dilated, hir_on_text_dilated = return_contours_of_image(text_regions_p_textregions_dilated)
|
||||
contours_only_dilated = return_parent_contours(contours_only_dilated, hir_on_text_dilated)
|
||||
|
||||
indexes_of_located_cont, center_x_coordinates_of_located, center_y_coordinates_of_located = self.return_indexes_of_contours_loctaed_inside_another_list_of_contours(contours_only_dilated, contours_only_text_parent_included, cx_conts_included, cy_conts_included, args_cont_located_included)
|
||||
|
||||
|
||||
if len(args_cont_located_excluded)>0:
|
||||
for ind in args_cont_located_excluded:
|
||||
indexes_of_located_cont.append(np.array([ind]))
|
||||
contours_only_dilated.append(contours_only_text_parent[ind])
|
||||
center_y_coordinates_of_located.append(0)
|
||||
|
||||
array_list = [np.array([elem]) if isinstance(elem, int) else elem for elem in indexes_of_located_cont]
|
||||
flattened_array = np.concatenate([arr.ravel() for arr in array_list])
|
||||
#print(len( np.unique(flattened_array)), 'indexes_of_located_cont uniques')
|
||||
|
||||
missing_textregions = list( set(np.array(range(len(contours_only_text_parent))) ) - set(np.unique(flattened_array)) )
|
||||
#print(missing_textregions, 'missing_textregions')
|
||||
|
||||
for ind in missing_textregions:
|
||||
indexes_of_located_cont.append(np.array([ind]))
|
||||
contours_only_dilated.append(contours_only_text_parent[ind])
|
||||
center_y_coordinates_of_located.append(0)
|
||||
|
||||
|
||||
if contours_only_text_parent_h:
|
||||
for vi in range(len(contours_only_text_parent_h)):
|
||||
indexes_of_located_cont.append(int(vi+len(contours_only_text_parent)))
|
||||
|
||||
array_list = [np.array([elem]) if isinstance(elem, int) else elem for elem in indexes_of_located_cont]
|
||||
flattened_array = np.concatenate([arr.ravel() for arr in array_list])
|
||||
|
||||
y_len = text_regions_p.shape[0]
|
||||
x_len = text_regions_p.shape[1]
|
||||
|
||||
img_poly = np.zeros((y_len,x_len), dtype='uint8')
|
||||
###img_poly[text_regions_p[:,:]==1] = 1
|
||||
###img_poly[text_regions_p[:,:]==2] = 2
|
||||
###img_poly[text_regions_p[:,:]==3] = 4
|
||||
###img_poly[text_regions_p[:,:]==6] = 5
|
||||
|
||||
##img_poly[text_regions_p[:,:]==1] = 1
|
||||
##img_poly[text_regions_p[:,:]==2] = 2
|
||||
##img_poly[text_regions_p[:,:]==3] = 3
|
||||
##img_poly[text_regions_p[:,:]==4] = 4
|
||||
##img_poly[text_regions_p[:,:]==5] = 5
|
||||
|
||||
img_poly = np.copy(text_regions_p)
|
||||
|
||||
img_header_and_sep = np.zeros((y_len,x_len), dtype='uint8')
|
||||
if contours_only_text_parent_h:
|
||||
_, cy_main, x_min_main, x_max_main, y_min_main, y_max_main, _ = find_new_features_of_contours(
|
||||
contours_only_text_parent_h)
|
||||
for j in range(len(cy_main)):
|
||||
img_header_and_sep[int(y_max_main[j]):int(y_max_main[j])+12,
|
||||
int(x_min_main[j]):int(x_max_main[j])] = 1
|
||||
co_text_all_org = contours_only_text_parent + contours_only_text_parent_h
|
||||
if len(contours_only_text_parent)>min_cont_size_to_be_dilated:
|
||||
co_text_all = contours_only_dilated + contours_only_text_parent_h
|
||||
else:
|
||||
co_text_all = contours_only_text_parent + contours_only_text_parent_h
|
||||
else:
|
||||
co_text_all_org = contours_only_text_parent
|
||||
if len(contours_only_text_parent)>min_cont_size_to_be_dilated:
|
||||
co_text_all = contours_only_dilated
|
||||
else:
|
||||
co_text_all = contours_only_text_parent
|
||||
|
||||
if not len(co_text_all):
|
||||
return [], []
|
||||
|
||||
labels_con = np.zeros((int(y_len /6.), int(x_len/6.), len(co_text_all)), dtype=bool)
|
||||
|
||||
co_text_all = [(i/6).astype(int) for i in co_text_all]
|
||||
for i in range(len(co_text_all)):
|
||||
img = labels_con[:,:,i].astype(np.uint8)
|
||||
|
||||
#img = cv2.resize(img, (int(img.shape[1]/6), int(img.shape[0]/6)), interpolation=cv2.INTER_NEAREST)
|
||||
|
||||
cv2.fillPoly(img, pts=[co_text_all[i]], color=(1,))
|
||||
labels_con[:,:,i] = img
|
||||
|
||||
|
||||
labels_con = resize_image(labels_con.astype(np.uint8), height1, width1).astype(bool)
|
||||
img_header_and_sep = resize_image(img_header_and_sep, height1, width1)
|
||||
img_poly = resize_image(img_poly, height3, width3)
|
||||
|
||||
|
||||
|
||||
input_1 = np.zeros((inference_bs, height1, width1, 3))
|
||||
ordered = [list(range(len(co_text_all)))]
|
||||
index_update = 0
|
||||
#print(labels_con.shape[2],"number of regions for reading order")
|
||||
while index_update>=0:
|
||||
ij_list = ordered.pop(index_update)
|
||||
i = ij_list.pop(0)
|
||||
|
||||
ante_list = []
|
||||
post_list = []
|
||||
tot_counter = 0
|
||||
batch = []
|
||||
for j in ij_list:
|
||||
img1 = labels_con[:,:,i].astype(float)
|
||||
img2 = labels_con[:,:,j].astype(float)
|
||||
img1[img_poly==5] = 2
|
||||
img2[img_poly==5] = 2
|
||||
img1[img_header_and_sep==1] = 3
|
||||
img2[img_header_and_sep==1] = 3
|
||||
|
||||
input_1[len(batch), :, :, 0] = img1 / 3.
|
||||
input_1[len(batch), :, :, 2] = img2 / 3.
|
||||
input_1[len(batch), :, :, 1] = img_poly / 5.
|
||||
|
||||
tot_counter += 1
|
||||
batch.append(j)
|
||||
if tot_counter % inference_bs == 0 or tot_counter == len(ij_list):
|
||||
y_pr = self.model_zoo.get('reading_order').predict(input_1 , verbose='0')
|
||||
for jb, j in enumerate(batch):
|
||||
if y_pr[jb][0]>=0.5:
|
||||
post_list.append(j)
|
||||
else:
|
||||
ante_list.append(j)
|
||||
batch = []
|
||||
|
||||
if len(ante_list):
|
||||
ordered.insert(index_update, ante_list)
|
||||
index_update += 1
|
||||
ordered.insert(index_update, [i])
|
||||
if len(post_list):
|
||||
ordered.insert(index_update + 1, post_list)
|
||||
|
||||
index_update = -1
|
||||
for index_next, ij_list in enumerate(ordered):
|
||||
if len(ij_list) > 1:
|
||||
index_update = index_next
|
||||
break
|
||||
|
||||
ordered = [i[0] for i in ordered]
|
||||
|
||||
##id_all_text = np.array(id_all_text)[index_sort]
|
||||
|
||||
|
||||
if len(contours_only_text_parent)>min_cont_size_to_be_dilated:
|
||||
org_contours_indexes = []
|
||||
for ind in range(len(ordered)):
|
||||
region_with_curr_order = ordered[ind]
|
||||
if region_with_curr_order < len(contours_only_dilated):
|
||||
if np.isscalar(indexes_of_located_cont[region_with_curr_order]):
|
||||
org_contours_indexes = org_contours_indexes + [indexes_of_located_cont[region_with_curr_order]]
|
||||
else:
|
||||
arg_sort_located_cont = np.argsort(center_y_coordinates_of_located[region_with_curr_order])
|
||||
org_contours_indexes = org_contours_indexes + list(np.array(indexes_of_located_cont[region_with_curr_order])[arg_sort_located_cont]) ##org_contours_indexes + list (
|
||||
else:
|
||||
org_contours_indexes = org_contours_indexes + [indexes_of_located_cont[region_with_curr_order]]
|
||||
|
||||
region_ids = ['region_%04d' % i for i in range(len(co_text_all_org))]
|
||||
return org_contours_indexes, region_ids
|
||||
else:
|
||||
region_ids = ['region_%04d' % i for i in range(len(co_text_all_org))]
|
||||
return ordered, region_ids
|
||||
|
||||
|
||||
|
||||
|
||||
def run(self,
|
||||
overwrite: bool = False,
|
||||
xml_filename: Optional[str] = None,
|
||||
dir_in: Optional[str] = None,
|
||||
dir_out: Optional[str] = None,
|
||||
):
|
||||
"""
|
||||
Get image and scales, then extract the page of scanned image
|
||||
"""
|
||||
self.logger.debug("enter run")
|
||||
t0_tot = time.time()
|
||||
|
||||
if dir_in:
|
||||
ls_xmls = [os.path.join(dir_in, xml_filename)
|
||||
for xml_filename in filter(is_xml_filename,
|
||||
os.listdir(dir_in))]
|
||||
elif xml_filename:
|
||||
ls_xmls = [xml_filename]
|
||||
else:
|
||||
raise ValueError("run requires either a single image filename or a directory")
|
||||
|
||||
for xml_filename in ls_xmls:
|
||||
self.logger.info(xml_filename)
|
||||
t0 = time.time()
|
||||
|
||||
file_name = Path(xml_filename).stem
|
||||
(tree_xml, root_xml, bb_coord_printspace, id_paragraph, id_header,
|
||||
co_text_paragraph, co_text_header, tot_region_ref,
|
||||
x_len, y_len, index_tot_regions, img_poly) = self.read_xml(xml_filename)
|
||||
|
||||
id_all_text = id_paragraph + id_header
|
||||
|
||||
order_text_new, id_of_texts_tot = self.do_order_of_regions_with_model(co_text_paragraph, co_text_header, img_poly[:,:,0])
|
||||
|
||||
id_all_text = np.array(id_all_text)[order_text_new]
|
||||
|
||||
alltags=[elem.tag for elem in root_xml.iter()]
|
||||
|
||||
|
||||
|
||||
link=alltags[0].split('}')[0]+'}'
|
||||
name_space = alltags[0].split('}')[0]
|
||||
name_space = name_space.split('{')[1]
|
||||
|
||||
page_element = root_xml.find(link+'Page')
|
||||
|
||||
|
||||
old_ro = root_xml.find(".//{*}ReadingOrder")
|
||||
|
||||
if old_ro is not None:
|
||||
page_element.remove(old_ro)
|
||||
|
||||
#print(old_ro, 'old_ro')
|
||||
ro_subelement = ET.Element('ReadingOrder')
|
||||
|
||||
ro_subelement2 = ET.SubElement(ro_subelement, 'OrderedGroup')
|
||||
ro_subelement2.set('id', "ro357564684568544579089")
|
||||
|
||||
for index, id_text in enumerate(id_all_text):
|
||||
new_element_2 = ET.SubElement(ro_subelement2, 'RegionRefIndexed')
|
||||
new_element_2.set('regionRef', id_all_text[index])
|
||||
new_element_2.set('index', str(index))
|
||||
|
||||
if (link+'PrintSpace' in alltags) or (link+'Border' in alltags):
|
||||
page_element.insert(1, ro_subelement)
|
||||
else:
|
||||
page_element.insert(0, ro_subelement)
|
||||
|
||||
alltags=[elem.tag for elem in root_xml.iter()]
|
||||
|
||||
ET.register_namespace("",name_space)
|
||||
assert dir_out
|
||||
tree_xml.write(os.path.join(dir_out, file_name+'.xml'),
|
||||
xml_declaration=True,
|
||||
method='xml',
|
||||
encoding="utf-8",
|
||||
default_namespace=None)
|
||||
|
||||
#sys.exit()
|
||||
|
||||
|
|
@ -1,4 +0,0 @@
|
|||
__all__ = [
|
||||
'EynollahModelZoo',
|
||||
]
|
||||
from .model_zoo import EynollahModelZoo
|
||||
|
|
@ -1,252 +0,0 @@
|
|||
from .specs import EynollahModelSpec, EynollahModelSpecSet
|
||||
|
||||
# NOTE: This needs to change whenever models/versions change
|
||||
ZENODO = "https://zenodo.org/records/17727267"
|
||||
MODELS_VERSION = "v0_8_0"
|
||||
|
||||
def dist_url(dist_name: str="layout") -> str:
|
||||
return f'{ZENODO}/models_{dist_name}_{MODELS_VERSION}.zip'
|
||||
|
||||
DEFAULT_MODEL_SPECS = EynollahModelSpecSet([
|
||||
|
||||
EynollahModelSpec(
|
||||
category="enhancement",
|
||||
variant='',
|
||||
filename="models_eynollah/eynollah-enhancement_20210425",
|
||||
dist_url=dist_url(),
|
||||
type='Keras',
|
||||
),
|
||||
|
||||
EynollahModelSpec(
|
||||
category="binarization",
|
||||
variant='hybrid',
|
||||
filename="models_eynollah/eynollah-binarization-hybrid_20230504/model_bin_hybrid_trans_cnn_sbb_ens",
|
||||
dist_url=dist_url(),
|
||||
type='Keras',
|
||||
),
|
||||
|
||||
EynollahModelSpec(
|
||||
category="binarization",
|
||||
variant='20210309',
|
||||
filename="models_eynollah/eynollah-binarization_20210309",
|
||||
dist_url=dist_url("extra"),
|
||||
type='Keras',
|
||||
),
|
||||
|
||||
EynollahModelSpec(
|
||||
category="binarization",
|
||||
variant='',
|
||||
filename="models_eynollah/eynollah-binarization_20210425",
|
||||
dist_url=dist_url("extra"),
|
||||
type='Keras',
|
||||
),
|
||||
|
||||
EynollahModelSpec(
|
||||
category="col_classifier",
|
||||
variant='',
|
||||
filename="models_eynollah/eynollah-column-classifier_20210425",
|
||||
dist_url=dist_url(),
|
||||
type='Keras',
|
||||
),
|
||||
|
||||
EynollahModelSpec(
|
||||
category="page",
|
||||
variant='',
|
||||
filename="models_eynollah/model_eynollah_page_extraction_20250915",
|
||||
dist_url=dist_url(),
|
||||
type='Keras',
|
||||
),
|
||||
|
||||
EynollahModelSpec(
|
||||
category="region",
|
||||
variant='',
|
||||
filename="models_eynollah/eynollah-main-regions-ensembled_20210425",
|
||||
dist_url=dist_url(),
|
||||
type='Keras',
|
||||
),
|
||||
|
||||
EynollahModelSpec(
|
||||
category="extract_images",
|
||||
variant='',
|
||||
filename="models_eynollah/eynollah-main-regions_20231127_672_org_ens_11_13_16_17_18",
|
||||
dist_url=dist_url(),
|
||||
type='Keras',
|
||||
),
|
||||
|
||||
EynollahModelSpec(
|
||||
category="region",
|
||||
variant='',
|
||||
filename="models_eynollah/eynollah-main-regions_20220314",
|
||||
dist_url=dist_url(),
|
||||
help="early layout",
|
||||
type='Keras',
|
||||
),
|
||||
|
||||
EynollahModelSpec(
|
||||
category="region_p2",
|
||||
variant='non-light',
|
||||
filename="models_eynollah/eynollah-main-regions-aug-rotation_20210425",
|
||||
dist_url=dist_url('extra'),
|
||||
help="early layout, non-light, 2nd part",
|
||||
type='Keras',
|
||||
),
|
||||
|
||||
EynollahModelSpec(
|
||||
category="region_1_2",
|
||||
variant='',
|
||||
#filename="models_eynollah/modelens_12sp_elay_0_3_4__3_6_n",
|
||||
#filename="models_eynollah/modelens_earlylayout_12spaltige_2_3_5_6_7_8",
|
||||
#filename="models_eynollah/modelens_early12_sp_2_3_5_6_7_8_9_10_12_14_15_16_18",
|
||||
#filename="models_eynollah/modelens_1_2_4_5_early_lay_1_2_spaltige",
|
||||
#filename="models_eynollah/model_3_eraly_layout_no_patches_1_2_spaltige",
|
||||
filename="models_eynollah/modelens_e_l_all_sp_0_1_2_3_4_171024",
|
||||
dist_url=dist_url("layout"),
|
||||
help="early layout, light, 1-or-2-column",
|
||||
type='Keras',
|
||||
),
|
||||
|
||||
EynollahModelSpec(
|
||||
category="region_fl_np",
|
||||
variant='',
|
||||
#'filename="models_eynollah/modelens_full_lay_1_3_031124",
|
||||
#'filename="models_eynollah/modelens_full_lay_13__3_19_241024",
|
||||
#'filename="models_eynollah/model_full_lay_13_241024",
|
||||
#'filename="models_eynollah/modelens_full_lay_13_17_231024",
|
||||
#'filename="models_eynollah/modelens_full_lay_1_2_221024",
|
||||
#'filename="models_eynollah/eynollah-full-regions-1column_20210425",
|
||||
filename="models_eynollah/modelens_full_lay_1__4_3_091124",
|
||||
dist_url=dist_url(),
|
||||
help="full layout / no patches",
|
||||
type='Keras',
|
||||
),
|
||||
|
||||
# FIXME: Why is region_fl and region_fl_np the same model?
|
||||
EynollahModelSpec(
|
||||
category="region_fl",
|
||||
variant='',
|
||||
# filename="models_eynollah/eynollah-full-regions-3+column_20210425",
|
||||
# filename="models_eynollah/model_2_full_layout_new_trans",
|
||||
# filename="models_eynollah/modelens_full_lay_1_3_031124",
|
||||
# filename="models_eynollah/modelens_full_lay_13__3_19_241024",
|
||||
# filename="models_eynollah/model_full_lay_13_241024",
|
||||
# filename="models_eynollah/modelens_full_lay_13_17_231024",
|
||||
# filename="models_eynollah/modelens_full_lay_1_2_221024",
|
||||
# filename="models_eynollah/modelens_full_layout_24_till_28",
|
||||
# filename="models_eynollah/model_2_full_layout_new_trans",
|
||||
filename="models_eynollah/modelens_full_lay_1__4_3_091124",
|
||||
dist_url=dist_url(),
|
||||
help="full layout / with patches",
|
||||
type='Keras',
|
||||
),
|
||||
|
||||
EynollahModelSpec(
|
||||
category="reading_order",
|
||||
variant='',
|
||||
#filename="models_eynollah/model_mb_ro_aug_ens_11",
|
||||
#filename="models_eynollah/model_step_3200000_mb_ro",
|
||||
#filename="models_eynollah/model_ens_reading_order_machine_based",
|
||||
#filename="models_eynollah/model_mb_ro_aug_ens_8",
|
||||
#filename="models_eynollah/model_ens_reading_order_machine_based",
|
||||
filename="models_eynollah/model_eynollah_reading_order_20250824",
|
||||
dist_url=dist_url(),
|
||||
type='Keras',
|
||||
),
|
||||
|
||||
EynollahModelSpec(
|
||||
category="textline",
|
||||
variant='non-light',
|
||||
#filename="models_eynollah/modelens_textline_1_4_16092024",
|
||||
#filename="models_eynollah/model_textline_ens_3_4_5_6_artificial",
|
||||
#filename="models_eynollah/modelens_textline_1_3_4_20240915",
|
||||
#filename="models_eynollah/model_textline_ens_3_4_5_6_artificial",
|
||||
#filename="models_eynollah/modelens_textline_9_12_13_14_15",
|
||||
#filename="models_eynollah/eynollah-textline_20210425",
|
||||
filename="models_eynollah/modelens_textline_0_1__2_4_16092024",
|
||||
dist_url=dist_url('extra'),
|
||||
type='Keras',
|
||||
),
|
||||
|
||||
EynollahModelSpec(
|
||||
category="textline",
|
||||
variant='',
|
||||
#filename="models_eynollah/eynollah-textline_light_20210425",
|
||||
filename="models_eynollah/modelens_textline_0_1__2_4_16092024",
|
||||
dist_url=dist_url(),
|
||||
type='Keras',
|
||||
),
|
||||
|
||||
EynollahModelSpec(
|
||||
category="table",
|
||||
variant='non-light',
|
||||
filename="models_eynollah/eynollah-tables_20210319",
|
||||
dist_url=dist_url('extra'),
|
||||
type='Keras',
|
||||
),
|
||||
|
||||
EynollahModelSpec(
|
||||
category="table",
|
||||
variant='',
|
||||
filename="models_eynollah/modelens_table_0t4_201124",
|
||||
dist_url=dist_url(),
|
||||
type='Keras',
|
||||
),
|
||||
|
||||
EynollahModelSpec(
|
||||
category="ocr",
|
||||
variant='',
|
||||
filename="models_eynollah/model_eynollah_ocr_cnnrnn_20250930",
|
||||
dist_url=dist_url("ocr"),
|
||||
type='Keras',
|
||||
),
|
||||
|
||||
EynollahModelSpec(
|
||||
category="ocr",
|
||||
variant='degraded',
|
||||
filename="models_eynollah/model_eynollah_ocr_cnnrnn__degraded_20250805/",
|
||||
help="slightly better at degraded Fraktur",
|
||||
dist_url=dist_url("ocr"),
|
||||
type='Keras',
|
||||
),
|
||||
|
||||
EynollahModelSpec(
|
||||
category="num_to_char",
|
||||
variant='',
|
||||
filename="characters_org.txt",
|
||||
dist_url=dist_url("ocr"),
|
||||
type='decoder',
|
||||
),
|
||||
|
||||
EynollahModelSpec(
|
||||
category="characters",
|
||||
variant='',
|
||||
filename="characters_org.txt",
|
||||
dist_url=dist_url("ocr"),
|
||||
type='List[str]',
|
||||
),
|
||||
|
||||
EynollahModelSpec(
|
||||
category="ocr",
|
||||
variant='tr',
|
||||
filename="models_eynollah/model_eynollah_ocr_trocr_20250919",
|
||||
dist_url=dist_url("ocr"),
|
||||
help='much slower transformer-based',
|
||||
type='Keras',
|
||||
),
|
||||
|
||||
EynollahModelSpec(
|
||||
category="trocr_processor",
|
||||
variant='',
|
||||
filename="models_eynollah/model_eynollah_ocr_trocr_20250919",
|
||||
dist_url=dist_url("ocr"),
|
||||
type='TrOCRProcessor',
|
||||
),
|
||||
|
||||
EynollahModelSpec(
|
||||
category="trocr_processor",
|
||||
variant='htr',
|
||||
filename="models_eynollah/microsoft/trocr-base-handwritten",
|
||||
dist_url=dist_url("extra"),
|
||||
type='TrOCRProcessor',
|
||||
),
|
||||
|
||||
])
|
||||
|
|
@ -1,281 +0,0 @@
|
|||
import os
|
||||
import json
|
||||
import logging
|
||||
from copy import deepcopy
|
||||
from pathlib import Path
|
||||
from fnmatch import fnmatchcase
|
||||
from typing import Dict, List, Optional, Tuple, Type, Union
|
||||
|
||||
from tabulate import tabulate
|
||||
|
||||
from ..predictor import Predictor
|
||||
from .specs import EynollahModelSpecSet
|
||||
from .default_specs import DEFAULT_MODEL_SPECS
|
||||
from .types import AnyModel, T
|
||||
|
||||
|
||||
class EynollahModelZoo:
|
||||
"""
|
||||
Wrapper class that handles storage and loading of models for all eynollah runners.
|
||||
"""
|
||||
|
||||
model_basedir: Path
|
||||
specs: EynollahModelSpecSet
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
basedir: str,
|
||||
model_overrides: Optional[List[Tuple[str, str, str]]] = None,
|
||||
) -> None:
|
||||
self.model_basedir = Path(basedir).resolve()
|
||||
self.logger = logging.getLogger('eynollah.model_zoo')
|
||||
if not self.model_basedir.exists():
|
||||
self.logger.warning(f"Model basedir does not exist: {basedir}. Set eynollah --model-basedir to the correct directory.")
|
||||
self.specs = deepcopy(DEFAULT_MODEL_SPECS)
|
||||
self._overrides = []
|
||||
if model_overrides:
|
||||
self.override_models(*model_overrides)
|
||||
self._loaded: Dict[str, Predictor] = {}
|
||||
|
||||
@property
|
||||
def model_overrides(self):
|
||||
return self._overrides
|
||||
|
||||
def override_models(
|
||||
self,
|
||||
*model_overrides: Tuple[str, str, str],
|
||||
):
|
||||
"""
|
||||
Override the default model versions
|
||||
"""
|
||||
for model_category, model_variant, model_filename in model_overrides:
|
||||
spec = self.specs.get(model_category, model_variant)
|
||||
self.logger.warning("Overriding filename for model spec %s to %s", spec, model_filename)
|
||||
self.specs.get(model_category, model_variant).filename = str(Path(model_filename).resolve())
|
||||
self._overrides += model_overrides
|
||||
|
||||
def model_path(
|
||||
self,
|
||||
model_category: str,
|
||||
model_variant: str = '',
|
||||
absolute: bool = True,
|
||||
) -> Path:
|
||||
"""
|
||||
Translate model_{type,variant} tuple into an absolute (or relative) Path
|
||||
"""
|
||||
spec = self.specs.get(model_category, model_variant)
|
||||
if spec.category in ('characters', 'num_to_char'):
|
||||
return self.model_path('ocr') / spec.filename
|
||||
if not Path(spec.filename).is_absolute() and absolute:
|
||||
model_path = Path(self.model_basedir).joinpath(spec.filename)
|
||||
else:
|
||||
model_path = Path(spec.filename)
|
||||
return model_path
|
||||
|
||||
def load_models(
|
||||
self,
|
||||
*all_load_args: Union[str, Tuple[str], Tuple[str, str], Tuple[str, str, str]],
|
||||
device: str = '',
|
||||
) -> Dict:
|
||||
"""
|
||||
Load all models by calling load_model and return a dictionary mapping model_category to loaded model
|
||||
"""
|
||||
ret = {} # cannot use self._loaded here, yet – first spawn all predictors
|
||||
for load_args in all_load_args:
|
||||
if isinstance(load_args, str):
|
||||
model_category = load_args
|
||||
load_args = [model_category]
|
||||
else:
|
||||
model_category = load_args[0]
|
||||
load_kwargs = {}
|
||||
if model_category.endswith('_resized'):
|
||||
load_args[0] = model_category[:-8]
|
||||
load_kwargs["resized"] = True
|
||||
elif model_category.endswith('_patched'):
|
||||
load_args[0] = model_category[:-8]
|
||||
load_kwargs["patched"] = True
|
||||
spec = self.specs.get(model_category, load_args[1] if len(load_args) > 1 else '')
|
||||
if spec.type in ['Keras'] and spec.category != 'ocr':
|
||||
ret[model_category] = Predictor(self.logger, self)
|
||||
ret[model_category].load_model(*load_args, **load_kwargs, device=device)
|
||||
else:
|
||||
ret[model_category] = self.load_model(*load_args, **load_kwargs, device=device)
|
||||
self._loaded.update(ret)
|
||||
return self._loaded
|
||||
|
||||
def load_model(
|
||||
self,
|
||||
model_category: str,
|
||||
model_variant: str = '',
|
||||
model_path_override: Optional[str] = None,
|
||||
patched: bool = False,
|
||||
resized: bool = False,
|
||||
device: str = '',
|
||||
) -> AnyModel:
|
||||
"""
|
||||
Load any model
|
||||
"""
|
||||
os.environ['TF_USE_LEGACY_KERAS'] = '1' # avoid Keras 3 after TF 2.15
|
||||
from ocrd_utils import tf_disable_interactive_logs
|
||||
tf_disable_interactive_logs()
|
||||
|
||||
import tensorflow as tf
|
||||
from tensorflow.keras.models import load_model
|
||||
|
||||
from ..patch_encoder import (
|
||||
PatchEncoder,
|
||||
Patches,
|
||||
wrap_layout_model_patched,
|
||||
wrap_layout_model_resized,
|
||||
)
|
||||
cuda = False
|
||||
try:
|
||||
gpus = tf.config.list_physical_devices('GPU')
|
||||
if device:
|
||||
if ',' in device:
|
||||
for spec in device.split(','):
|
||||
cat, dev = spec.split(':')
|
||||
if fnmatchcase(model_category, cat):
|
||||
device = dev
|
||||
break
|
||||
if device == 'CPU':
|
||||
gpus = []
|
||||
else:
|
||||
assert device.startswith('GPU')
|
||||
gpus = [gpus[int(device[3:])]]
|
||||
else:
|
||||
gpus = gpus[:1] # TF will always use first allowable
|
||||
tf.config.set_visible_devices(gpus, 'GPU')
|
||||
for device in gpus:
|
||||
tf.config.experimental.set_memory_growth(device, True)
|
||||
vendor_name = (
|
||||
tf.config.experimental.get_device_details(device)
|
||||
.get('device_name', 'unknown'))
|
||||
cuda = True
|
||||
self.logger.info("using GPU %s (%s) for model %s",
|
||||
device.name,
|
||||
vendor_name,
|
||||
model_category + (
|
||||
"_patched" if patched else
|
||||
"_resized" if resized else ""))
|
||||
except RuntimeError:
|
||||
self.logger.exception("cannot configure GPU devices")
|
||||
if not cuda:
|
||||
self.logger.warning("no GPU device available")
|
||||
|
||||
if model_path_override:
|
||||
self.override_models((model_category, model_variant, model_path_override))
|
||||
model_path = self.model_path(model_category, model_variant)
|
||||
if model_path.suffix == '.h5' and Path(model_path.stem).exists():
|
||||
# prefer SavedModel over HDF5 format if it exists
|
||||
model_path = Path(model_path.stem)
|
||||
if model_category == 'ocr':
|
||||
model = self._load_ocr_model(variant=model_variant)
|
||||
elif model_category == 'num_to_char':
|
||||
model = self._load_num_to_char()
|
||||
elif model_category == 'characters':
|
||||
model = self._load_characters()
|
||||
elif model_category == 'trocr_processor':
|
||||
from transformers import TrOCRProcessor
|
||||
model = TrOCRProcessor.from_pretrained(model_path)
|
||||
else:
|
||||
try:
|
||||
# avoid wasting VRAM on non-transformer models
|
||||
model = load_model(model_path, compile=False)
|
||||
except Exception as e:
|
||||
self.logger.error(e)
|
||||
model = load_model(
|
||||
model_path, compile=False,
|
||||
custom_objects=dict(PatchEncoder=PatchEncoder,
|
||||
Patches=Patches))
|
||||
model._name = model_category
|
||||
if resized:
|
||||
model = wrap_layout_model_resized(model)
|
||||
model._name = model_category + '_resized'
|
||||
elif patched:
|
||||
model = wrap_layout_model_patched(model)
|
||||
model._name = model_category + '_patched'
|
||||
else:
|
||||
model.jit_compile = True
|
||||
model.make_predict_function()
|
||||
return model
|
||||
|
||||
def get(self, model_category: str) -> Predictor:
|
||||
if model_category not in self._loaded:
|
||||
raise ValueError(f'Model "{model_category}" not previously loaded with "load_model(..)"')
|
||||
return self._loaded[model_category]
|
||||
|
||||
def _load_ocr_model(self, variant: str) -> AnyModel:
|
||||
"""
|
||||
Load OCR model
|
||||
"""
|
||||
from tensorflow.keras.models import Model as KerasModel
|
||||
from tensorflow.keras.models import load_model
|
||||
|
||||
ocr_model_dir = self.model_path('ocr', variant)
|
||||
if variant == 'tr':
|
||||
from transformers import VisionEncoderDecoderModel
|
||||
ret = VisionEncoderDecoderModel.from_pretrained(ocr_model_dir)
|
||||
assert isinstance(ret, VisionEncoderDecoderModel)
|
||||
return ret
|
||||
else:
|
||||
ocr_model = load_model(ocr_model_dir, compile=False)
|
||||
assert isinstance(ocr_model, KerasModel)
|
||||
return KerasModel(
|
||||
ocr_model.get_layer(name="image").input, # type: ignore
|
||||
ocr_model.get_layer(name="dense2").output, # type: ignore
|
||||
)
|
||||
|
||||
def _load_characters(self) -> List[str]:
|
||||
"""
|
||||
Load encoding for OCR
|
||||
"""
|
||||
with open(self.model_path('num_to_char'), "r") as config_file:
|
||||
return json.load(config_file)
|
||||
|
||||
def _load_num_to_char(self) -> 'StringLookup':
|
||||
"""
|
||||
Load decoder for OCR
|
||||
"""
|
||||
from tensorflow.keras.layers import StringLookup
|
||||
|
||||
characters = self._load_characters()
|
||||
# Mapping characters to integers.
|
||||
char_to_num = StringLookup(vocabulary=characters, mask_token=None)
|
||||
# Mapping integers back to original characters.
|
||||
return StringLookup(vocabulary=char_to_num.get_vocabulary(), mask_token=None, invert=True)
|
||||
|
||||
def __str__(self):
|
||||
return tabulate(
|
||||
[
|
||||
[
|
||||
spec.type,
|
||||
spec.category,
|
||||
spec.variant,
|
||||
spec.help,
|
||||
f'Yes, at {self.model_path(spec.category, spec.variant)}'
|
||||
if self.model_path(spec.category, spec.variant).exists()
|
||||
else f'No, download {spec.dist_url}',
|
||||
# self.model_path(spec.category, spec.variant),
|
||||
]
|
||||
for spec in sorted(self.specs.specs, key=lambda x: x.dist_url)
|
||||
],
|
||||
headers=[
|
||||
'Type',
|
||||
'Category',
|
||||
'Variant',
|
||||
'Help',
|
||||
'Used in',
|
||||
'Installed',
|
||||
],
|
||||
tablefmt='github',
|
||||
)
|
||||
|
||||
def shutdown(self):
|
||||
"""
|
||||
Ensure that a loaded models is not referenced by ``self._loaded`` anymore
|
||||
"""
|
||||
if hasattr(self, '_loaded') and getattr(self, '_loaded'):
|
||||
for needle in list(self._loaded.keys()):
|
||||
self._loaded[needle].shutdown()
|
||||
del self._loaded[needle]
|
||||
|
|
@ -1,52 +0,0 @@
|
|||
from dataclasses import dataclass
|
||||
from typing import Dict, List, Set, Tuple
|
||||
|
||||
|
||||
@dataclass
|
||||
class EynollahModelSpec():
|
||||
"""
|
||||
Describing a single model abstractly.
|
||||
"""
|
||||
category: str
|
||||
# Relative filename to the models_eynollah directory in the dists
|
||||
filename: str
|
||||
# URL to the smallest model distribution containing this model (link to Zenodo)
|
||||
dist_url: str
|
||||
type: str
|
||||
variant: str = ''
|
||||
help: str = ''
|
||||
|
||||
class EynollahModelSpecSet():
|
||||
"""
|
||||
List of all used models for eynollah.
|
||||
"""
|
||||
specs: List[EynollahModelSpec]
|
||||
|
||||
def __init__(self, specs: List[EynollahModelSpec]) -> None:
|
||||
self.specs = sorted(specs, key=lambda x: x.category + '0' + x.variant)
|
||||
self.categories: Set[str] = set([spec.category for spec in self.specs])
|
||||
self.variants: Dict[str, Set[str]] = {
|
||||
spec.category: set([x.variant for x in self.specs if x.category == spec.category])
|
||||
for spec in self.specs
|
||||
}
|
||||
self._index_category_variant: Dict[Tuple[str, str], EynollahModelSpec] = {
|
||||
(spec.category, spec.variant): spec
|
||||
for spec in self.specs
|
||||
}
|
||||
|
||||
def asdict(self) -> Dict[str, Dict[str, str]]:
|
||||
return {
|
||||
spec.category: {
|
||||
spec.variant: spec.filename
|
||||
}
|
||||
for spec in self.specs
|
||||
}
|
||||
|
||||
def get(self, category: str, variant: str) -> EynollahModelSpec:
|
||||
if category not in self.categories:
|
||||
raise ValueError(f"Unknown category '{category}', must be one of {self.categories}")
|
||||
if variant not in self.variants[category]:
|
||||
raise ValueError(f"Unknown variant {variant} for {category}. Known variants: {self.variants[category]}")
|
||||
return self._index_category_variant[(category, variant)]
|
||||
|
||||
|
||||
|
|
@ -1,7 +0,0 @@
|
|||
from typing import TypeVar
|
||||
|
||||
# NOTE: Creating an actual union type requires loading transformers which is expensive and error-prone
|
||||
# from transformers import TrOCRProcessor, VisionEncoderDecoderModel
|
||||
# AnyModel = Union[VisionEncoderDecoderModel, TrOCRProcessor, KerasModel, List]
|
||||
AnyModel = object
|
||||
T = TypeVar('T')
|
||||
|
|
@ -1,183 +0,0 @@
|
|||
{
|
||||
"version": "0.8.0",
|
||||
"git_url": "https://github.com/qurator-spk/eynollah",
|
||||
"dockerhub": "ocrd/eynollah",
|
||||
"tools": {
|
||||
"ocrd-eynollah-segment": {
|
||||
"executable": "ocrd-eynollah-segment",
|
||||
"categories": ["Layout analysis"],
|
||||
"description": "Segment page into regions and lines and do reading order detection with eynollah",
|
||||
"input_file_grp_cardinality": 1,
|
||||
"output_file_grp_cardinality": 1,
|
||||
"steps": ["layout/segmentation/region", "layout/segmentation/line"],
|
||||
"parameters": {
|
||||
"models": {
|
||||
"type": "string",
|
||||
"format": "uri",
|
||||
"content-type": "text/directory",
|
||||
"cacheable": true,
|
||||
"description": "Directory containing models to be used (See https://qurator-data.de/eynollah)",
|
||||
"required": true
|
||||
},
|
||||
"dpi": {
|
||||
"type": "number",
|
||||
"format": "float",
|
||||
"description": "pixel density in dots per inch (overrides any meta-data in the images); ignored if <= 0 (with fall-back 230)",
|
||||
"default": 0
|
||||
},
|
||||
"full_layout": {
|
||||
"type": "boolean",
|
||||
"default": true,
|
||||
"description": "Try to detect all region subtypes, including drop-capital and heading"
|
||||
},
|
||||
"light_version": {
|
||||
"type": "boolean",
|
||||
"default": true,
|
||||
"enum": [true],
|
||||
"description": "ignored (only for backwards-compatibility)"
|
||||
},
|
||||
"textline_light": {
|
||||
"type": "boolean",
|
||||
"default": true,
|
||||
"enum": [true],
|
||||
"description": "ignored (only for backwards-compatibility)"
|
||||
},
|
||||
"tables": {
|
||||
"type": "boolean",
|
||||
"default": false,
|
||||
"description": "Try to detect table regions"
|
||||
},
|
||||
"curved_line": {
|
||||
"type": "boolean",
|
||||
"default": false,
|
||||
"description": "retrieve textline polygons independent of each other (needs more processing time)"
|
||||
},
|
||||
"ignore_page_extraction": {
|
||||
"type": "boolean",
|
||||
"default": false,
|
||||
"description": "if true, do not attempt page frame detection (cropping)"
|
||||
},
|
||||
"allow_scaling": {
|
||||
"type": "boolean",
|
||||
"default": false,
|
||||
"description": "check the resolution against the number of detected columns and if needed, scale the image up or down during layout detection (heuristic to improve quality and performance)"
|
||||
},
|
||||
"allow_enhancement": {
|
||||
"type": "boolean",
|
||||
"default": false,
|
||||
"description": "if this parameter set to true, this tool would check that input image need resizing and enhancement or not."
|
||||
},
|
||||
"right_to_left": {
|
||||
"type": "boolean",
|
||||
"default": false,
|
||||
"description": "if true, return reading order in right-to-left reading direction."
|
||||
},
|
||||
"headers_off": {
|
||||
"type": "boolean",
|
||||
"default": false,
|
||||
"description": "ignore the special role of headings during reading order detection"
|
||||
},
|
||||
"reading_order_machine_based": {
|
||||
"type": "boolean",
|
||||
"default": false,
|
||||
"description": "use data-driven (rather than rule-based) reading order detection"
|
||||
}
|
||||
},
|
||||
"resources": [
|
||||
{
|
||||
"url": "https://zenodo.org/records/17727267/files/models_all_v0_8_0.zip",
|
||||
"name": "models_all_v0_8_0",
|
||||
"type": "archive",
|
||||
"size": 5636009377,
|
||||
"description": "Models for layout detection, reading order detection, textline detection, page extraction, column classification, table detection, binarization and image enhancement",
|
||||
"version_range": ">= v0.8.0"
|
||||
},
|
||||
{
|
||||
"url": "https://zenodo.org/records/17580627/files/models_all_v0_7_0.zip?download=1",
|
||||
"name": "models_layout_v0_7_0",
|
||||
"type": "archive",
|
||||
"size": 6119874002,
|
||||
"description": "Models for layout detection, reading order detection, textline detection, page extraction, column classification, table detection, binarization, image enhancement and OCR",
|
||||
"version_range": ">= v0.7.0"
|
||||
},
|
||||
{
|
||||
"url": "https://zenodo.org/records/17295988/files/models_layout_v0_6_0.tar.gz?download=1",
|
||||
"name": "models_layout_v0_6_0",
|
||||
"type": "archive",
|
||||
"path_in_archive": "models_layout_v0_6_0",
|
||||
"size": 3525684179,
|
||||
"description": "Models for layout detection, reading order detection, textline detection, page extraction, column classification, table detection, binarization, image enhancement and OCR",
|
||||
"version_range": ">= v0.5.0"
|
||||
},
|
||||
{
|
||||
"description": "models for eynollah (TensorFlow SavedModel format)",
|
||||
"url": "https://github.com/qurator-spk/eynollah/releases/download/v0.3.1/models_eynollah.tar.gz",
|
||||
"name": "default",
|
||||
"size": 1894627041,
|
||||
"type": "archive",
|
||||
"path_in_archive": "models_eynollah",
|
||||
"version_range": ">= v0.3.0, < v0.5.0"
|
||||
}
|
||||
]
|
||||
},
|
||||
"ocrd-sbb-binarize": {
|
||||
"executable": "ocrd-sbb-binarize",
|
||||
"description": "Pixelwise binarization with selectional auto-encoders in Keras",
|
||||
"categories": ["Image preprocessing"],
|
||||
"steps": ["preprocessing/optimization/binarization"],
|
||||
"input_file_grp_cardinality": 1,
|
||||
"output_file_grp_cardinality": 1,
|
||||
"parameters": {
|
||||
"operation_level": {
|
||||
"type": "string",
|
||||
"enum": ["page", "region"],
|
||||
"default": "page",
|
||||
"description": "PAGE XML hierarchy level to operate on"
|
||||
},
|
||||
"model": {
|
||||
"description": "Directory containing HDF5 or SavedModel/ProtoBuf models. Can be an absolute path or a path relative to the OCR-D resource location, the current working directory or the $SBB_BINARIZE_DATA environment variable (if set)",
|
||||
"type": "string",
|
||||
"format": "uri",
|
||||
"content-type": "text/directory",
|
||||
"required": true
|
||||
}
|
||||
},
|
||||
"resources": [
|
||||
{
|
||||
"url": "https://zenodo.org/records/17727267/files/models_all_v0_8_0.zip",
|
||||
"name": "models_all_v0_8_0",
|
||||
"type": "archive",
|
||||
"size": 5636009377,
|
||||
"description": "Models for layout detection, reading order detection, textline detection, page extraction, column classification, table detection, binarization and image enhancement",
|
||||
"version_range": ">= v0.8.0"
|
||||
},
|
||||
{
|
||||
"url": "https://zenodo.org/records/17580627/files/models_all_v0_7_0.zip?download=1",
|
||||
"name": "models_layout_v0_7_0",
|
||||
"type": "archive",
|
||||
"size": 6119874002,
|
||||
"description": "Models for layout detection, reading order detection, textline detection, page extraction, column classification, table detection, binarization, image enhancement and OCR",
|
||||
"version_range": ">= v0.7.0"
|
||||
},
|
||||
{
|
||||
"url": "https://github.com/qurator-spk/sbb_binarization/releases/download/v0.0.11/saved_model_2020_01_16.zip",
|
||||
"name": "default",
|
||||
"type": "archive",
|
||||
"path_in_archive": "saved_model_2020_01_16",
|
||||
"size": 563147331,
|
||||
"description": "default models provided by github.com/qurator-spk (SavedModel format)",
|
||||
"version_range": "< v0.7.0"
|
||||
},
|
||||
{
|
||||
"url": "https://github.com/qurator-spk/sbb_binarization/releases/download/v0.0.11/saved_model_2021_03_09.zip",
|
||||
"name": "default-2021-03-09",
|
||||
"type": "archive",
|
||||
"path_in_archive": ".",
|
||||
"size": 133230419,
|
||||
"description": "updated default models provided by github.com/qurator-spk (SavedModel format)",
|
||||
"version_range": "< v0.7.0"
|
||||
}
|
||||
]
|
||||
}
|
||||
}
|
||||
}
|
||||
|
|
@ -1,109 +0,0 @@
|
|||
from functools import cached_property
|
||||
from typing import Optional
|
||||
|
||||
from PIL import Image
|
||||
from frozendict import frozendict
|
||||
import numpy as np
|
||||
import cv2
|
||||
from click import command
|
||||
|
||||
from ocrd import Processor, OcrdPageResult, OcrdPageResultImage
|
||||
from ocrd_models.ocrd_page import OcrdPage, AlternativeImageType
|
||||
from ocrd.decorators import ocrd_cli_options, ocrd_cli_wrap_processor
|
||||
|
||||
from eynollah.model_zoo.model_zoo import EynollahModelZoo
|
||||
|
||||
from .sbb_binarize import SbbBinarizer
|
||||
from .utils.pil_cv2 import cv2pil
|
||||
|
||||
|
||||
class SbbBinarizeProcessor(Processor):
|
||||
# already employs GPU (without singleton process atm)
|
||||
max_workers = 1
|
||||
|
||||
@cached_property
|
||||
def executable(self):
|
||||
return 'ocrd-sbb-binarize'
|
||||
|
||||
def setup(self):
|
||||
"""
|
||||
Set up the model prior to processing.
|
||||
"""
|
||||
# resolve relative path via OCR-D ResourceManager
|
||||
assert isinstance(self.parameter, frozendict)
|
||||
model_zoo = EynollahModelZoo(basedir=self.parameter['model'])
|
||||
self.binarizer = SbbBinarizer(model_zoo=model_zoo, logger=self.logger)
|
||||
|
||||
def process_page_pcgts(self, *input_pcgts: Optional[OcrdPage], page_id: Optional[str] = None) -> OcrdPageResult:
|
||||
"""
|
||||
Binarize images with sbb_binarization (based on selectional auto-encoders).
|
||||
|
||||
For each page of the input file group, open and deserialize input PAGE-XML
|
||||
and its respective images. Then iterate over the element hierarchy down to
|
||||
the requested ``operation_level``.
|
||||
|
||||
For each segment element, retrieve a raw (non-binarized) segment image
|
||||
according to the layout annotation (from an existing ``AlternativeImage``,
|
||||
or by cropping into the higher-level images, and deskewing when applicable).
|
||||
|
||||
Pass the image to the binarizer (which runs in fixed-size windows/patches
|
||||
across the image and stitches the results together).
|
||||
|
||||
Serialize the resulting bilevel image as PNG file and add it to the output
|
||||
file group (with file ID suffix ``.IMG-BIN``) along with the output PAGE-XML
|
||||
(referencing it as new ``AlternativeImage`` for the segment element).
|
||||
|
||||
Produce a new PAGE output file by serialising the resulting hierarchy.
|
||||
"""
|
||||
assert input_pcgts
|
||||
assert input_pcgts[0]
|
||||
assert self.parameter
|
||||
oplevel = self.parameter['operation_level']
|
||||
pcgts = input_pcgts[0]
|
||||
result = OcrdPageResult(pcgts)
|
||||
page = pcgts.get_Page()
|
||||
page_image, page_xywh, _ = self.workspace.image_from_page(
|
||||
page, page_id, feature_filter='binarized')
|
||||
|
||||
if oplevel == 'page':
|
||||
self.logger.info("Binarizing on 'page' level in page '%s'", page_id)
|
||||
page_image_bin = cv2pil(self.binarizer.run_single("", img_pil=page_image,
|
||||
use_patches=True))
|
||||
# update PAGE (reference the image file):
|
||||
page_image_ref = AlternativeImageType(comments=page_xywh['features'] + ',binarized,clipped')
|
||||
page.add_AlternativeImage(page_image_ref)
|
||||
result.images.append(OcrdPageResultImage(page_image_bin, '.IMG-BIN', page_image_ref))
|
||||
|
||||
elif oplevel == 'region':
|
||||
regions = page.get_AllRegions(['Text', 'Table'], depth=1)
|
||||
if not regions:
|
||||
self.logger.warning("Page '%s' contains no text/table regions", page_id)
|
||||
for region in regions:
|
||||
region_image, region_xywh = self.workspace.image_from_segment(
|
||||
region, page_image, page_xywh, feature_filter='binarized')
|
||||
region_image_bin = cv2pil(self.binarizer.run_single("", img_pil=region_image,
|
||||
use_patches=True))
|
||||
# update PAGE (reference the image file):
|
||||
region_image_ref = AlternativeImageType(comments=region_xywh['features'] + ',binarized')
|
||||
region.add_AlternativeImage(region_image_ref)
|
||||
result.images.append(OcrdPageResultImage(region_image_bin, region.id + '.IMG-BIN', region_image_ref))
|
||||
|
||||
elif oplevel == 'line':
|
||||
lines = page.get_AllTextLines()
|
||||
if not lines:
|
||||
self.logger.warning("Page '%s' contains no text lines", page_id)
|
||||
for line in lines:
|
||||
line_image, line_xywh = self.workspace.image_from_segment(line, page_image, page_xywh, feature_filter='binarized')
|
||||
line_image_bin = cv2pil(self.binarizer.run_single("", img_pil=line_image,
|
||||
use_patches=True))
|
||||
# update PAGE (reference the image file):
|
||||
line_image_ref = AlternativeImageType(comments=line_xywh['features'] + ',binarized')
|
||||
line.add_AlternativeImage(line_image_ref)
|
||||
result.images.append(OcrdPageResultImage(line_image_bin, line.id + '.IMG-BIN', line_image_ref))
|
||||
|
||||
return result
|
||||
|
||||
@command()
|
||||
@ocrd_cli_options
|
||||
def main(*args, **kwargs):
|
||||
return ocrd_cli_wrap_processor(SbbBinarizeProcessor, *args, **kwargs)
|
||||
|
|
@ -1,156 +0,0 @@
|
|||
import os
|
||||
os.environ['TF_USE_LEGACY_KERAS'] = '1' # avoid Keras 3 after TF 2.15
|
||||
import tensorflow as tf
|
||||
from tensorflow.keras import layers, models
|
||||
|
||||
class PatchEncoder(layers.Layer):
|
||||
|
||||
# 441=21*21 # 14*14 # 28*28
|
||||
def __init__(self, num_patches=441, projection_dim=64):
|
||||
super().__init__()
|
||||
self.num_patches = num_patches
|
||||
self.projection_dim = projection_dim
|
||||
self.projection = layers.Dense(self.projection_dim)
|
||||
self.position_embedding = layers.Embedding(self.num_patches, self.projection_dim)
|
||||
|
||||
def call(self, patch):
|
||||
positions = tf.range(start=0, limit=self.num_patches, delta=1)
|
||||
return self.projection(patch) + self.position_embedding(positions)
|
||||
|
||||
def get_config(self):
|
||||
return dict(num_patches=self.num_patches,
|
||||
projection_dim=self.projection_dim,
|
||||
**super().get_config())
|
||||
|
||||
class Patches(layers.Layer):
|
||||
def __init__(self, patch_size_x=1, patch_size_y=1):
|
||||
super().__init__()
|
||||
self.patch_size_x = patch_size_x
|
||||
self.patch_size_y = patch_size_y
|
||||
|
||||
def call(self, images):
|
||||
batch_size = tf.shape(images)[0]
|
||||
patches = tf.image.extract_patches(
|
||||
images=images,
|
||||
sizes=[1, self.patch_size_y, self.patch_size_x, 1],
|
||||
strides=[1, self.patch_size_y, self.patch_size_x, 1],
|
||||
rates=[1, 1, 1, 1],
|
||||
padding="VALID",
|
||||
)
|
||||
patch_dims = patches.shape[-1]
|
||||
return tf.reshape(patches, [batch_size, -1, patch_dims])
|
||||
|
||||
def get_config(self):
|
||||
return dict(patch_size_x=self.patch_size_x,
|
||||
patch_size_y=self.patch_size_y,
|
||||
**super().get_config())
|
||||
|
||||
class wrap_layout_model_resized(models.Model):
|
||||
"""
|
||||
replacement for layout model using resizing to model width/height and back
|
||||
|
||||
(accepts arbitrary width/height input [B, H, W, 3], returns same size segmentation [B, H, W, C])
|
||||
"""
|
||||
def __init__(self, model):
|
||||
super().__init__(name=model.name + '_resized')
|
||||
self.model = model
|
||||
self.height = model.layers[-1].output_shape[1]
|
||||
self.width = model.layers[-1].output_shape[2]
|
||||
|
||||
@tf.function(reduce_retracing=True,
|
||||
#jit_compile=True, (ScaleAndTranslate is not supported by XLA)
|
||||
input_signature=[tf.TensorSpec([1, None, None, 3],
|
||||
dtype=tf.float32)])
|
||||
def call(self, img, training=False):
|
||||
height = tf.shape(img)[1]
|
||||
width = tf.shape(img)[2]
|
||||
img_resized = tf.image.resize(img,
|
||||
(self.height, self.width),
|
||||
antialias=True)
|
||||
pred_resized = self.model(img_resized)
|
||||
pred = tf.image.resize(pred_resized,
|
||||
(height, width))
|
||||
return pred
|
||||
|
||||
class wrap_layout_model_patched(models.Model):
|
||||
"""
|
||||
replacement for layout model using sliding window for patches
|
||||
|
||||
(accepts arbitrary width/height input [B, H, W, 3], returns same size segmentation [B, H, W, C])
|
||||
"""
|
||||
def __init__(self, model):
|
||||
super().__init__(name=model.name + '_patched')
|
||||
self.model = model
|
||||
self.height = model.layers[-1].output_shape[1]
|
||||
self.width = model.layers[-1].output_shape[2]
|
||||
self.classes = model.layers[-1].output_shape[3]
|
||||
# equivalent of marginal_of_patch_percent=0.1 ...
|
||||
self.stride_x = int(self.width * (1 - 0.1))
|
||||
self.stride_y = int(self.height * (1 - 0.1))
|
||||
offset_height = (self.height - self.stride_y) // 2
|
||||
offset_width = (self.width - self.stride_x) // 2
|
||||
window = tf.image.pad_to_bounding_box(
|
||||
tf.ones((self.stride_y, self.stride_x, 1), dtype=tf.int32),
|
||||
offset_height, offset_width,
|
||||
self.height, self.width)
|
||||
self.window = tf.expand_dims(window, axis=0)
|
||||
|
||||
@tf.function(reduce_retracing=True,
|
||||
#jit_compile=True, (ScaleAndTranslate and ExtractImagePatches not supported by XLA)
|
||||
input_signature=[tf.TensorSpec([1, None, None, 3],
|
||||
dtype=tf.float32)])
|
||||
def call(self, img, training=False):
|
||||
height = tf.shape(img)[1]
|
||||
width = tf.shape(img)[2]
|
||||
if (height < self.height or
|
||||
width < self.width):
|
||||
img_resized = tf.image.resize(img,
|
||||
(self.height, self.width),
|
||||
antialias=True)
|
||||
pred_resized = self.model(img_resized)
|
||||
pred = tf.image.resize(pred_resized,
|
||||
(height, width))
|
||||
return pred
|
||||
|
||||
img_patches = tf.image.extract_patches(
|
||||
images=img,
|
||||
sizes=[1, self.height, self.width, 1],
|
||||
strides=[1, self.stride_y, self.stride_x, 1],
|
||||
rates=[1, 1, 1, 1],
|
||||
padding='SAME')
|
||||
img_patches = tf.squeeze(img_patches)
|
||||
index_shape = (-1, self.height, self.width, 2)
|
||||
input_shape = (-1, self.height, self.width, 3)
|
||||
output_shape = (-1, self.height, self.width, self.classes)
|
||||
img_patches = tf.reshape(img_patches, shape=input_shape)
|
||||
# may be too large:
|
||||
#pred_patches = self.model(img_patches)
|
||||
# so rebatch to fit in memory:
|
||||
img_patches = tf.expand_dims(img_patches, 1)
|
||||
pred_patches = tf.map_fn(self.model, img_patches,
|
||||
parallel_iterations=1,
|
||||
infer_shape=False)
|
||||
pred_patches = tf.squeeze(pred_patches, 1)
|
||||
# calculate corresponding indexes for reconstruction
|
||||
x = tf.range(width)
|
||||
y = tf.range(height)
|
||||
x, y = tf.meshgrid(x, y)
|
||||
indices = tf.stack([y, x], axis=-1)
|
||||
indices_patches = tf.image.extract_patches(
|
||||
images=tf.expand_dims(indices, axis=0),
|
||||
sizes=[1, self.height, self.width, 1],
|
||||
strides=[1, self.stride_y, self.stride_x, 1],
|
||||
rates=[1, 1, 1, 1],
|
||||
padding='SAME')
|
||||
indices_patches = tf.squeeze(indices_patches)
|
||||
indices_patches = tf.reshape(indices_patches, shape=index_shape)
|
||||
|
||||
# use margins for sliding window approach
|
||||
indices_patches = indices_patches * self.window
|
||||
|
||||
pred = tf.scatter_nd(
|
||||
indices_patches,
|
||||
pred_patches,
|
||||
(height, width, self.classes))
|
||||
pred = tf.expand_dims(pred, axis=0)
|
||||
return pred
|
||||
|
|
@ -1,210 +0,0 @@
|
|||
from contextlib import ExitStack
|
||||
from typing import List, Dict
|
||||
import logging
|
||||
import logging.handlers
|
||||
import multiprocessing as mp
|
||||
import numpy as np
|
||||
|
||||
from .utils.shm import share_ndarray, ndarray_shared
|
||||
|
||||
QSIZE = 200
|
||||
|
||||
|
||||
class Predictor(mp.context.SpawnProcess):
|
||||
"""
|
||||
singleton subprocess solely responsible for prediction with TensorFlow,
|
||||
communicates with any number of worker processes,
|
||||
acting as a shallow replacement for various model types in EynollahModelZoo's
|
||||
_loaded dict for each single model
|
||||
"""
|
||||
def __init__(self, logger, model_zoo):
|
||||
self.logger = logger
|
||||
self.loglevel = logger.parent.level
|
||||
self.model_zoo = model_zoo
|
||||
ctxt = mp.get_context('spawn')
|
||||
self.taskq = ctxt.Queue(maxsize=QSIZE)
|
||||
self.resultq = ctxt.Queue(maxsize=QSIZE)
|
||||
self.logq = ctxt.Queue(maxsize=QSIZE * 100)
|
||||
logging.handlers.QueueListener(
|
||||
self.logq, *(
|
||||
# as per ocrd_utils.initLogging():
|
||||
logging.root.handlers +
|
||||
# as per eynollah_cli.main():
|
||||
self.logger.parent.handlers
|
||||
), respect_handler_level=False).start()
|
||||
self.stopped = ctxt.Event()
|
||||
self.closable = ctxt.Manager().list()
|
||||
super().__init__(name="EynollahPredictor", daemon=True)
|
||||
|
||||
@property
|
||||
def input_shape(self):
|
||||
return self({})
|
||||
|
||||
def predict(self, data: dict, verbose=0):
|
||||
return self(data)
|
||||
|
||||
def __call__(self, data: dict):
|
||||
# unusable as per python/cpython#79967
|
||||
#with self.jobid.get_lock():
|
||||
# would work, but not public:
|
||||
#with self.jobid._mutex:
|
||||
with self.joblock:
|
||||
self.jobid.value += 1
|
||||
jobid = self.jobid.value
|
||||
if not len(data):
|
||||
self.taskq.put((jobid, data))
|
||||
#self.logger.debug("sent shape query task '%d' for model '%s'", jobid, self.name)
|
||||
return self.result(jobid)
|
||||
with share_ndarray(data) as shared_data:
|
||||
self.taskq.put((jobid, shared_data))
|
||||
#self.logger.debug("sent prediction task '%d' for model '%s': %s", jobid, self.name, shared_data)
|
||||
return self.result(jobid)
|
||||
|
||||
def result(self, jobid):
|
||||
while not self.stopped.is_set():
|
||||
if jobid in self.results:
|
||||
#self.logger.debug("received result for '%d'", jobid)
|
||||
result = self.results.pop(jobid)
|
||||
if isinstance(result, Exception):
|
||||
raise Exception(f"predictor {self.name} failed for {jobid}") from result
|
||||
elif isinstance(result, dict):
|
||||
with ndarray_shared(result) as shared_result:
|
||||
result = np.copy(shared_result)
|
||||
self.closable.append(jobid)
|
||||
return result
|
||||
try:
|
||||
jobid0, result = self.resultq.get(timeout=0.7)
|
||||
except mp.queues.Empty:
|
||||
continue
|
||||
#self.logger.debug("storing results for '%d': '%s'", jobid0, result)
|
||||
self.results[jobid0] = result
|
||||
raise Exception(f"predictor {self.name} terminated while waiting on results for {jobid}")
|
||||
|
||||
def run(self):
|
||||
try:
|
||||
self.setup() # fill model_zoo etc
|
||||
except Exception as e:
|
||||
self.logger.exception("setup failed")
|
||||
self.stopped.set()
|
||||
return
|
||||
closing = {}
|
||||
def close_all():
|
||||
for jobid in list(self.closable):
|
||||
self.closable.remove(jobid)
|
||||
closing.pop(jobid).close()
|
||||
#self.logger.debug("closed shm for '%d'", jobid)
|
||||
while not self.stopped.is_set():
|
||||
close_all()
|
||||
try:
|
||||
TIMEOUT = 4.5 # 1.1 too is greedy - not enough for rebatching
|
||||
jobid, shared_data = self.taskq.get(timeout=TIMEOUT)
|
||||
except mp.queues.Empty:
|
||||
continue
|
||||
try:
|
||||
# up to what batch size fits into small (8GB) VRAM?
|
||||
# (notice we are not listing _resized/_patched models here,
|
||||
# because its inputs/outputs will have varying shapes)
|
||||
REBATCH_SIZE = {
|
||||
# small models (448x448)...
|
||||
"col_classifier": 2,
|
||||
"page": 2,
|
||||
"binarization": 4,
|
||||
"enhancement": 4,
|
||||
"reading_order": 4,
|
||||
# medium size (672x672x3)...
|
||||
"textline": 2,
|
||||
# large models...
|
||||
"table": 1,
|
||||
"region_1_2": 1,
|
||||
"region_fl_np": 1,
|
||||
"region_fl": 1,
|
||||
}.get(self.name, 1)
|
||||
REBATCH_SIZE = 1 # save VRAM; FIXME: re-enable w/ runtime parameter
|
||||
if not len(shared_data):
|
||||
#self.logger.debug("getting '%d' output shape of model '%s'", jobid, self.name)
|
||||
result = self.model.input_shape
|
||||
self.resultq.put((jobid, result))
|
||||
#self.logger.debug("sent result for '%d': %s", jobid, result)
|
||||
else:
|
||||
tasks = [(jobid, shared_data)]
|
||||
batch_size = shared_data['shape'][0]
|
||||
while (not self.taskq.empty() and
|
||||
# climb to target batch size
|
||||
batch_size * len(tasks) < REBATCH_SIZE):
|
||||
jobid0, shared_data0 = self.taskq.get()
|
||||
if len(shared_data0):
|
||||
# add to our batch
|
||||
tasks.append((jobid0, shared_data0))
|
||||
else:
|
||||
# immediately anser
|
||||
self.resultq.put((jobid0, self.model.input_shape))
|
||||
if len(tasks) > 1:
|
||||
self.logger.debug("rebatching %d '%s' tasks of batch size %d",
|
||||
len(tasks), self.name, batch_size)
|
||||
with ExitStack() as stack:
|
||||
data = []
|
||||
jobs = []
|
||||
for jobid, shared_data in tasks:
|
||||
#self.logger.debug("predicting '%d' with model '%s': %s", jobid, self.name, shared_data)
|
||||
jobs.append(jobid)
|
||||
data.append(stack.enter_context(ndarray_shared(shared_data)))
|
||||
data = np.concatenate(data)
|
||||
#result = self.model.predict(data, verbose=0)
|
||||
# faster, less VRAM
|
||||
result = self.model.predict_on_batch(data)
|
||||
results = np.split(result, len(jobs))
|
||||
#self.logger.debug("sharing result array for '%d'", jobid)
|
||||
with ExitStack() as stack:
|
||||
for jobid, result in zip(jobs, results):
|
||||
# we don't know when the result will be received,
|
||||
# but don't want to wait either, so track closing
|
||||
# context per job, and wait for closable signal
|
||||
# from client
|
||||
result = stack.enter_context(share_ndarray(result))
|
||||
closing[jobid] = stack.pop_all()
|
||||
self.resultq.put((jobid, result))
|
||||
#self.logger.debug("sent result for '%d': %s", jobid, result)
|
||||
except Exception as e:
|
||||
self.logger.error("prediction for %s failed: %s", self.name, e.__class__.__name__)
|
||||
result = e
|
||||
self.resultq.put((jobid, result))
|
||||
close_all()
|
||||
#self.logger.debug("predictor terminated")
|
||||
|
||||
def load_model(self, *load_args, **load_kwargs):
|
||||
assert len(load_args)
|
||||
self.name = '_'.join(list(load_args[:1]) +
|
||||
list(key for key in load_kwargs
|
||||
if key != 'device'))
|
||||
self.load_args = load_args
|
||||
self.load_kwargs = load_kwargs
|
||||
self.start() # call run() in subprocess
|
||||
# parent context here
|
||||
del self.model_zoo # only in subprocess
|
||||
ctxt = mp.get_context('fork') # ocrd.Processor will fork workers
|
||||
mngr = ctxt.Manager()
|
||||
self.jobid = mngr.Value('i', 0)
|
||||
self.joblock = mngr.Lock()
|
||||
self.results = mngr.dict()
|
||||
|
||||
def setup(self):
|
||||
logging.root.handlers = [logging.handlers.QueueHandler(self.logq)]
|
||||
self.logger.setLevel(self.loglevel)
|
||||
self.model = self.model_zoo.load_model(*self.load_args, **self.load_kwargs)
|
||||
|
||||
def shutdown(self):
|
||||
# do not terminate from forked processor instances
|
||||
if mp.parent_process() is None:
|
||||
self.stopped.set()
|
||||
self.taskq.close()
|
||||
self.taskq.cancel_join_thread()
|
||||
self.resultq.close()
|
||||
self.resultq.cancel_join_thread()
|
||||
self.logq.close()
|
||||
self.terminate()
|
||||
else:
|
||||
del self.model
|
||||
|
||||
def __del__(self):
|
||||
#self.logger.debug(f"deinit of {self} in {mp.current_process().name}")
|
||||
self.shutdown()
|
||||
|
|
@ -1,83 +0,0 @@
|
|||
from functools import cached_property
|
||||
from typing import Optional
|
||||
from ocrd_models import OcrdPage
|
||||
from ocrd import OcrdPageResultImage, Processor, OcrdPageResult
|
||||
|
||||
from eynollah.model_zoo.model_zoo import EynollahModelZoo
|
||||
|
||||
from .eynollah import Eynollah, EynollahXmlWriter
|
||||
|
||||
class EynollahProcessor(Processor):
|
||||
@cached_property
|
||||
def executable(self) -> str:
|
||||
return 'ocrd-eynollah-segment'
|
||||
|
||||
def setup(self) -> None:
|
||||
assert self.parameter
|
||||
model_zoo = EynollahModelZoo(basedir=self.parameter['models'])
|
||||
self.eynollah = Eynollah(
|
||||
model_zoo=model_zoo,
|
||||
allow_enhancement=self.parameter['allow_enhancement'],
|
||||
curved_line=self.parameter['curved_line'],
|
||||
right2left=self.parameter['right_to_left'],
|
||||
reading_order_machine_based=self.parameter['reading_order_machine_based'],
|
||||
ignore_page_extraction=self.parameter['ignore_page_extraction'],
|
||||
full_layout=self.parameter['full_layout'],
|
||||
allow_scaling=self.parameter['allow_scaling'],
|
||||
headers_off=self.parameter['headers_off'],
|
||||
tables=self.parameter['tables'],
|
||||
logger=self.logger
|
||||
)
|
||||
self.eynollah.plotter = None
|
||||
|
||||
def shutdown(self):
|
||||
if hasattr(self, 'eynollah'):
|
||||
del self.eynollah
|
||||
|
||||
def process_page_pcgts(self, *input_pcgts: Optional[OcrdPage], page_id: Optional[str] = None) -> OcrdPageResult:
|
||||
"""
|
||||
Performs cropping, region and line segmentation with Eynollah.
|
||||
|
||||
For each page, open and deserialize PAGE input file (from existing
|
||||
PAGE file in the input fileGrp, or generated from image file).
|
||||
Retrieve its respective page-level image (ignoring annotation that
|
||||
already added `binarized`, `cropped` or `deskewed` features).
|
||||
|
||||
Set up Eynollah to detect regions and lines, and add each one to the
|
||||
page, respectively.
|
||||
|
||||
\b
|
||||
- If ``tables``, try to detect table blocks and add them as TableRegion.
|
||||
- If ``full_layout``, then in addition to paragraphs and marginals, also
|
||||
try to detect drop capitals and headings.
|
||||
- If ``ignore_page_extraction``, then attempt no cropping of the page.
|
||||
- If ``curved_line``, then compute contour polygons for text lines
|
||||
instead of simple bounding boxes.
|
||||
- If ``reading_order_machine_based``, then detect reading order via
|
||||
data-driven model instead of geometrical heuristics.
|
||||
|
||||
Produce a new output file by serialising the resulting hierarchy.
|
||||
"""
|
||||
assert input_pcgts
|
||||
assert input_pcgts[0]
|
||||
assert self.parameter
|
||||
pcgts = input_pcgts[0]
|
||||
result = OcrdPageResult(pcgts)
|
||||
page = pcgts.get_Page()
|
||||
page_image, _, _ = self.workspace.image_from_page(
|
||||
page, page_id,
|
||||
# avoid any features that would change the coordinate system: cropped,deskewed
|
||||
# (the PAGE builder merely adds regions, so afterwards we would not know which to transform)
|
||||
# also avoid binarization as models usually fare better on grayscale/RGB
|
||||
feature_filter='cropped,deskewed,binarized')
|
||||
if hasattr(page_image, 'filename'):
|
||||
image_filename = page_image.filename
|
||||
else:
|
||||
image_filename = "dummy" # will be replaced by ocrd.Processor.process_page_file
|
||||
result.images.append(OcrdPageResultImage(page_image, '.IMG', page)) # mark as new original
|
||||
# FIXME: mask out already existing regions (incremental segmentation)
|
||||
self.eynollah.run_single(image_filename,
|
||||
img_pil=page_image, pcgts=pcgts,
|
||||
# ocrd.Processor will handle OCRD_EXISTING_OUTPUT more flexibly
|
||||
overwrite=True)
|
||||
return result
|
||||
|
|
@ -1,91 +0,0 @@
|
|||
"""
|
||||
Tool to load model and binarize a given image.
|
||||
"""
|
||||
|
||||
# pyright: reportIndexIssue=false
|
||||
# pyright: reportCallIssue=false
|
||||
# pyright: reportArgumentType=false
|
||||
# pyright: reportPossiblyUnboundVariable=false
|
||||
|
||||
import os
|
||||
import logging
|
||||
from pathlib import Path
|
||||
from typing import Optional
|
||||
|
||||
import numpy as np
|
||||
import cv2
|
||||
|
||||
from .eynollah import Eynollah
|
||||
from .model_zoo import EynollahModelZoo
|
||||
from .utils.resize import resize_image
|
||||
from .utils import is_image_filename
|
||||
|
||||
class SbbBinarizer(Eynollah):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
*,
|
||||
model_zoo: EynollahModelZoo,
|
||||
logger: Optional[logging.Logger] = None,
|
||||
device: str = '',
|
||||
):
|
||||
self.logger = logger if logger else logging.getLogger('eynollah.binarization')
|
||||
self.model_zoo = model_zoo
|
||||
self.setup_models(device=device)
|
||||
|
||||
def setup_models(self, device=''):
|
||||
loadable = ['binarization']
|
||||
self.model_zoo.load_models(*loadable, device=device)
|
||||
for model in loadable:
|
||||
self.logger.debug("model %s has input shape %s", model,
|
||||
self.model_zoo.get(model).input_shape)
|
||||
|
||||
def run(self,
|
||||
image=None,
|
||||
image_filename=None,
|
||||
output=None,
|
||||
use_patches=False,
|
||||
dir_in=None,
|
||||
overwrite=False
|
||||
):
|
||||
"""
|
||||
Binarize the scanned images
|
||||
"""
|
||||
if dir_in:
|
||||
ls_imgs = [(os.path.join(dir_in, image_filename),
|
||||
os.path.join(output, Path(image_filename).stem + '.png'))
|
||||
for image_filename in filter(is_image_filename,
|
||||
os.listdir(dir_in))]
|
||||
elif image_filename:
|
||||
ls_imgs = [(image_filename, output)]
|
||||
else:
|
||||
raise ValueError("run requires either a single image filename or a directory")
|
||||
|
||||
for img_filename, output_filename in ls_imgs:
|
||||
self.logger.info(img_filename)
|
||||
|
||||
if os.path.exists(output_filename):
|
||||
if overwrite:
|
||||
self.logger.warning("will overwrite existing output file '%s'", output_filename)
|
||||
else:
|
||||
self.logger.warning("will skip input for existing output file '%s'", output_filename)
|
||||
continue
|
||||
|
||||
img_res = self.run_single(img_filename,
|
||||
use_patches=use_patches)
|
||||
|
||||
cv2.imwrite(output_filename, img_res)
|
||||
self.logger.info("output filename: '%s'", output_filename)
|
||||
|
||||
def run_single(self,
|
||||
img_filename: str,
|
||||
img_pil=None,
|
||||
use_patches: bool = False,
|
||||
):
|
||||
image = self.cache_images(image_filename=img_filename, image_pil=img_pil)
|
||||
img = self.imread(image)
|
||||
img_bin = self.do_prediction(use_patches, img, self.model_zoo.get("binarization"),
|
||||
n_batch_inference=5)
|
||||
img_bin = 255 * (img_bin == 0).astype(np.uint8)
|
||||
#img_bin = np.repeat(img_bin[:, :, np.newaxis], 3, axis=2).astype(np.uint8)
|
||||
return img_bin
|
||||
|
|
@ -1,18 +0,0 @@
|
|||
import sys
|
||||
import click
|
||||
|
||||
from .models import resnet50_unet
|
||||
|
||||
|
||||
@click.command()
|
||||
def build_model_load_pretrained_weights_and_save():
|
||||
n_classes = 2
|
||||
input_height = 224
|
||||
input_width = 448
|
||||
weight_decay = 1e-6
|
||||
pretraining = False
|
||||
dir_of_weights = 'model_bin_sbb_ens.h5'
|
||||
|
||||
model = resnet50_unet(n_classes, input_height, input_width, weight_decay, pretraining)
|
||||
model.load_weights(dir_of_weights)
|
||||
model.save('./name_in_another_python_version.h5')
|
||||
|
|
@ -1,30 +0,0 @@
|
|||
import os
|
||||
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
|
||||
|
||||
import click
|
||||
import sys
|
||||
|
||||
from .build_model_load_pretrained_weights_and_save import build_model_load_pretrained_weights_and_save
|
||||
from .generate_gt_for_training import main as generate_gt_cli
|
||||
from .inference import main as inference_cli
|
||||
from .train import ex
|
||||
from .extract_line_gt import linegt_cli
|
||||
from .weights_ensembling import ensemble_cli
|
||||
|
||||
@click.command(context_settings=dict(
|
||||
ignore_unknown_options=True,
|
||||
))
|
||||
@click.argument('SACRED_ARGS', nargs=-1, type=click.UNPROCESSED)
|
||||
def train_cli(sacred_args):
|
||||
ex.run_commandline([sys.argv[0]] + list(sacred_args))
|
||||
|
||||
@click.group('training')
|
||||
def main():
|
||||
pass
|
||||
|
||||
main.add_command(build_model_load_pretrained_weights_and_save)
|
||||
main.add_command(generate_gt_cli, 'generate-gt')
|
||||
main.add_command(inference_cli, 'inference')
|
||||
main.add_command(train_cli, 'train')
|
||||
main.add_command(linegt_cli, 'export_textline_images_and_text')
|
||||
main.add_command(ensemble_cli, 'ensembling')
|
||||
|
|
@ -1,134 +0,0 @@
|
|||
from logging import Logger, getLogger
|
||||
from typing import Optional
|
||||
from pathlib import Path
|
||||
import os
|
||||
|
||||
import click
|
||||
import cv2
|
||||
import xml.etree.ElementTree as ET
|
||||
import numpy as np
|
||||
|
||||
from ..utils import is_image_filename
|
||||
|
||||
@click.command()
|
||||
@click.option(
|
||||
"--image",
|
||||
"-i",
|
||||
help="input image filename",
|
||||
type=click.Path(exists=True, dir_okay=False),
|
||||
)
|
||||
@click.option(
|
||||
"--dir_in",
|
||||
"-di",
|
||||
help="directory of input images (instead of --image)",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
)
|
||||
@click.option(
|
||||
"--dir_xmls",
|
||||
"-dx",
|
||||
help="directory of input PAGE-XML files (in addition to --dir_in; filename stems must match the image files, with '.xml' suffix).",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
required=True,
|
||||
)
|
||||
@click.option(
|
||||
"--out",
|
||||
"-o",
|
||||
'dir_out',
|
||||
help="directory for output PAGE-XML files",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
required=True,
|
||||
)
|
||||
@click.option(
|
||||
"--dataset_abbrevation",
|
||||
"-ds_pref",
|
||||
'pref_of_dataset',
|
||||
help="in the case of extracting textline and text from a xml GT file user can add an abbrevation of dataset name to generated dataset",
|
||||
)
|
||||
@click.option(
|
||||
"--do_not_mask_with_textline_contour",
|
||||
"-nmtc/-mtc",
|
||||
is_flag=True,
|
||||
help="if this parameter set to true, cropped textline images will not be masked with textline contour.",
|
||||
)
|
||||
def linegt_cli(
|
||||
image,
|
||||
dir_in,
|
||||
dir_xmls,
|
||||
dir_out,
|
||||
pref_of_dataset,
|
||||
do_not_mask_with_textline_contour,
|
||||
):
|
||||
assert bool(dir_in) ^ bool(image), "Set --dir-in or --image-filename, not both"
|
||||
if dir_in:
|
||||
ls_imgs = [
|
||||
os.path.join(dir_in, image) for image in filter(is_image_filename, os.listdir(dir_in))
|
||||
]
|
||||
else:
|
||||
assert image
|
||||
ls_imgs = [image]
|
||||
|
||||
for dir_img in ls_imgs:
|
||||
file_name = Path(dir_img).stem
|
||||
dir_xml = os.path.join(dir_xmls, file_name + '.xml')
|
||||
|
||||
img = cv2.imread(dir_img)
|
||||
|
||||
total_bb_coordinates = []
|
||||
|
||||
tree1 = ET.parse(dir_xml, parser=ET.XMLParser(encoding="utf-8"))
|
||||
root1 = tree1.getroot()
|
||||
alltags = [elem.tag for elem in root1.iter()]
|
||||
|
||||
name_space = alltags[0].split('}')[0]
|
||||
name_space = name_space.split('{')[1]
|
||||
|
||||
region_tags = np.unique([x for x in alltags if x.endswith('TextRegion')])
|
||||
|
||||
cropped_lines_region_indexer = []
|
||||
|
||||
indexer_text_region = 0
|
||||
indexer_textlines = 0
|
||||
# FIXME: non recursive, use OCR-D PAGE generateDS API. Or use an existing tool for this purpose altogether
|
||||
for nn in root1.iter(region_tags):
|
||||
for child_textregion in nn:
|
||||
if child_textregion.tag.endswith("TextLine"):
|
||||
for child_textlines in child_textregion:
|
||||
if child_textlines.tag.endswith("Coords"):
|
||||
cropped_lines_region_indexer.append(indexer_text_region)
|
||||
p_h = child_textlines.attrib['points'].split(' ')
|
||||
textline_coords = np.array([[int(x.split(',')[0]), int(x.split(',')[1])] for x in p_h])
|
||||
|
||||
x, y, w, h = cv2.boundingRect(textline_coords)
|
||||
|
||||
total_bb_coordinates.append([x, y, w, h])
|
||||
|
||||
img_poly_on_img = np.copy(img)
|
||||
|
||||
mask_poly = np.zeros(img.shape)
|
||||
mask_poly = cv2.fillPoly(mask_poly, pts=[textline_coords], color=(1, 1, 1))
|
||||
|
||||
mask_poly = mask_poly[y : y + h, x : x + w, :]
|
||||
img_crop = img_poly_on_img[y : y + h, x : x + w, :]
|
||||
|
||||
if not do_not_mask_with_textline_contour:
|
||||
img_crop[mask_poly == 0] = 255
|
||||
|
||||
if img_crop.shape[0] == 0 or img_crop.shape[1] == 0:
|
||||
continue
|
||||
if child_textlines.tag.endswith("TextEquiv"):
|
||||
for cheild_text in child_textlines:
|
||||
if cheild_text.tag.endswith("Unicode"):
|
||||
textline_text = cheild_text.text
|
||||
if textline_text:
|
||||
base_name = os.path.join(
|
||||
dir_out, file_name + '_line_' + str(indexer_textlines)
|
||||
)
|
||||
if pref_of_dataset:
|
||||
base_name += '_' + pref_of_dataset
|
||||
if not do_not_mask_with_textline_contour:
|
||||
base_name += '_masked'
|
||||
|
||||
with open(base_name + '.txt', 'w') as text_file:
|
||||
text_file.write(textline_text)
|
||||
cv2.imwrite(base_name + '.png', img_crop)
|
||||
indexer_textlines += 1
|
||||
|
|
@ -1,621 +0,0 @@
|
|||
import click
|
||||
import json
|
||||
import os
|
||||
from tqdm import tqdm
|
||||
from pathlib import Path
|
||||
from PIL import Image, ImageDraw, ImageFont
|
||||
import cv2
|
||||
import numpy as np
|
||||
|
||||
from .gt_gen_utils import (
|
||||
filter_contours_area_of_image,
|
||||
find_format_of_given_filename_in_dir,
|
||||
find_new_features_of_contours,
|
||||
fit_text_single_line,
|
||||
get_content_of_dir,
|
||||
get_images_of_ground_truth,
|
||||
get_layout_contours_for_visualization,
|
||||
get_textline_contours_and_ocr_text,
|
||||
get_textline_contours_for_visualization,
|
||||
overlay_layout_on_image,
|
||||
read_xml,
|
||||
resize_image,
|
||||
visualize_image_from_contours,
|
||||
visualize_image_from_contours_layout
|
||||
)
|
||||
|
||||
@click.group()
|
||||
def main():
|
||||
"""
|
||||
extract GT data suitable for model training for various tasks
|
||||
"""
|
||||
pass
|
||||
|
||||
@main.command()
|
||||
@click.option(
|
||||
"--dir_xml",
|
||||
"-dx",
|
||||
help="input directory of GT PAGE-XML files",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
required=True,
|
||||
)
|
||||
@click.option(
|
||||
"--dir_images",
|
||||
"-di",
|
||||
help="input directory of GT image files (only needed for '--printspace' or scaling configured via 'columns_width'; filename stems should match those in --dir_xml)",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
)
|
||||
@click.option(
|
||||
"--dir_out_images",
|
||||
"-doi",
|
||||
help="output directory for training image files (for printspace cropping or scaling)",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
)
|
||||
@click.option(
|
||||
"--dir_out",
|
||||
"-do",
|
||||
help="output directory for training label files",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
required=True,
|
||||
)
|
||||
|
||||
@click.option(
|
||||
"--config",
|
||||
"-cfg",
|
||||
help="config file of prefered layout or use case.",
|
||||
type=click.Path(exists=True, dir_okay=False),
|
||||
)
|
||||
|
||||
@click.option(
|
||||
"--type_output",
|
||||
"-to",
|
||||
type=click.Choice(["2d", "3d"]),
|
||||
default="2d",
|
||||
help="generate labels as [H, W] array pseudo index-color images for training ('2d') or [H, W, C] array RGB color images for plotting ('3d')",
|
||||
)
|
||||
@click.option(
|
||||
"--printspace",
|
||||
"-ps",
|
||||
is_flag=True,
|
||||
help="crop pages from annotated PrintSpace or Border to generate labels and images (will also require -di for so original images so output images are cropped along with labels)",
|
||||
)
|
||||
@click.option(
|
||||
"--missing-printspace",
|
||||
"-mps",
|
||||
type=click.Choice(["full", "skip", "project"]),
|
||||
default="full",
|
||||
help="if -ps is set, what to do in case a PAGE-XML has no PrintSpace or Border annotation: keep entire page ('full'), ignore file ('skip') or crop artificially from outer hull of all segments ('project')",
|
||||
)
|
||||
|
||||
def pagexml2label(dir_xml, dir_out, type_output, config, printspace, missing_printspace, dir_images, dir_out_images):
|
||||
"""
|
||||
extract PAGE-XML GT data suitable for model training for segmentation tasks
|
||||
"""
|
||||
if config:
|
||||
with open(config) as f:
|
||||
config_params = json.load(f)
|
||||
else:
|
||||
print("passed")
|
||||
config_params = None
|
||||
get_images_of_ground_truth(get_content_of_dir(dir_xml),
|
||||
dir_xml,
|
||||
dir_out,
|
||||
type_output,
|
||||
config,
|
||||
config_params,
|
||||
printspace,
|
||||
missing_printspace,
|
||||
dir_images,
|
||||
dir_out_images
|
||||
)
|
||||
|
||||
@main.command()
|
||||
@click.option(
|
||||
"--dir_imgs",
|
||||
"-dis",
|
||||
help="directory of images with high resolution.",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
)
|
||||
@click.option(
|
||||
"--dir_out_images",
|
||||
"-dois",
|
||||
help="directory where degraded images will be written.",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
)
|
||||
|
||||
@click.option(
|
||||
"--dir_out_labels",
|
||||
"-dols",
|
||||
help="directory where original images will be written as labels.",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
)
|
||||
@click.option(
|
||||
"--scales",
|
||||
"-scs",
|
||||
help="json dictionary where the scales are written.",
|
||||
type=click.Path(exists=True, dir_okay=False),
|
||||
)
|
||||
def image_enhancement(dir_imgs, dir_out_images, dir_out_labels, scales):
|
||||
"""
|
||||
extract image GT data suitable for model training for image enhancement tasks
|
||||
"""
|
||||
ls_imgs = os.listdir(dir_imgs)
|
||||
with open(scales) as f:
|
||||
scale_dict = json.load(f)
|
||||
ls_scales = scale_dict['scales']
|
||||
|
||||
for img in tqdm(ls_imgs):
|
||||
img_name = img.split('.')[0]
|
||||
img_type = img.split('.')[1]
|
||||
image = cv2.imread(os.path.join(dir_imgs, img))
|
||||
for i, scale in enumerate(ls_scales):
|
||||
height_sc = int(image.shape[0]*scale)
|
||||
width_sc = int(image.shape[1]*scale)
|
||||
|
||||
image_down_scaled = resize_image(image, height_sc, width_sc)
|
||||
image_back_to_org_scale = resize_image(image_down_scaled, image.shape[0], image.shape[1])
|
||||
|
||||
cv2.imwrite(os.path.join(dir_out_images, img_name+'_'+str(i)+'.'+img_type), image_back_to_org_scale)
|
||||
cv2.imwrite(os.path.join(dir_out_labels, img_name+'_'+str(i)+'.'+img_type), image)
|
||||
|
||||
|
||||
@main.command()
|
||||
@click.option(
|
||||
"--dir_xml",
|
||||
"-dx",
|
||||
help="directory of GT page-xml files",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
)
|
||||
|
||||
@click.option(
|
||||
"--dir_out_modal_image",
|
||||
"-domi",
|
||||
help="directory where ground truth images would be written",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
)
|
||||
|
||||
@click.option(
|
||||
"--dir_out_classes",
|
||||
"-docl",
|
||||
help="directory where ground truth classes would be written",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
)
|
||||
|
||||
@click.option(
|
||||
"--input_height",
|
||||
"-ih",
|
||||
help="input height",
|
||||
)
|
||||
@click.option(
|
||||
"--input_width",
|
||||
"-iw",
|
||||
help="input width",
|
||||
)
|
||||
@click.option(
|
||||
"--min_area_size",
|
||||
"-min",
|
||||
help="min area size of regions considered for reading order training.",
|
||||
)
|
||||
|
||||
@click.option(
|
||||
"--min_area_early",
|
||||
"-min_early",
|
||||
help="If you have already generated a training dataset using a specific minimum area value and now wish to create a dataset with a smaller minimum area value, you can avoid regenerating the previous dataset by providing the earlier minimum area value. This will ensure that only the missing data is generated.",
|
||||
)
|
||||
|
||||
def machine_based_reading_order(dir_xml, dir_out_modal_image, dir_out_classes, input_height, input_width, min_area_size, min_area_early):
|
||||
"""
|
||||
extract PAGE-XML GT data suitable for model training for reading-order task
|
||||
"""
|
||||
xml_files_ind = os.listdir(dir_xml)
|
||||
xml_files_ind = [ind_xml for ind_xml in xml_files_ind if ind_xml.endswith('.xml')]
|
||||
input_height = int(input_height)
|
||||
input_width = int(input_width)
|
||||
min_area = float(min_area_size)
|
||||
if min_area_early:
|
||||
min_area_early = float(min_area_early)
|
||||
|
||||
|
||||
indexer_start= 0#55166
|
||||
max_area = 1
|
||||
#min_area = 0.0001
|
||||
|
||||
for ind_xml in tqdm(xml_files_ind):
|
||||
indexer = 0
|
||||
#print(ind_xml)
|
||||
#print('########################')
|
||||
xml_file = os.path.join(dir_xml,ind_xml )
|
||||
f_name = ind_xml.split('.')[0]
|
||||
_, _, _, file_name, id_paragraph, id_header,co_text_paragraph,co_text_header,tot_region_ref,x_len, y_len,index_tot_regions,img_poly = read_xml(xml_file)
|
||||
|
||||
id_all_text = id_paragraph + id_header
|
||||
co_text_all = co_text_paragraph + co_text_header
|
||||
|
||||
|
||||
_, cy_main, x_min_main, x_max_main, y_min_main, y_max_main, _ = find_new_features_of_contours(co_text_header)
|
||||
|
||||
img_header_and_sep = np.zeros((y_len,x_len), dtype='uint8')
|
||||
|
||||
for j in range(len(cy_main)):
|
||||
img_header_and_sep[int(y_max_main[j]):int(y_max_main[j])+12,
|
||||
int(x_min_main[j]):int(x_max_main[j]) ] = 1
|
||||
|
||||
|
||||
try:
|
||||
texts_corr_order_index_int = [int(index_tot_regions[tot_region_ref.index(i)])
|
||||
for i in id_all_text]
|
||||
except ValueError as e:
|
||||
print("incomplete ReadingOrder in", xml_file, "- skipping:", str(e))
|
||||
continue
|
||||
|
||||
co_text_all, texts_corr_order_index_int, regions_ar_less_than_early_min = \
|
||||
filter_contours_area_of_image(img_poly, co_text_all, texts_corr_order_index_int,
|
||||
max_area, min_area, min_area_early)
|
||||
|
||||
|
||||
arg_array = np.array(range(len(texts_corr_order_index_int)))
|
||||
|
||||
labels_con = np.zeros((y_len,x_len,len(arg_array)),dtype='uint8')
|
||||
for i in range(len(co_text_all)):
|
||||
img_label = np.zeros((y_len,x_len,3),dtype='uint8')
|
||||
img_label=cv2.fillPoly(img_label, pts =[co_text_all[i]], color=(1,1,1))
|
||||
|
||||
img_label[:,:,0][img_poly[:,:,0]==5] = 2
|
||||
img_label[:,:,0][img_header_and_sep[:,:]==1] = 3
|
||||
|
||||
labels_con[:,:,i] = img_label[:,:,0]
|
||||
|
||||
labels_con = resize_image(labels_con, input_height, input_width)
|
||||
img_poly = resize_image(img_poly, input_height, input_width)
|
||||
|
||||
|
||||
for i in range(len(texts_corr_order_index_int)):
|
||||
for j in range(len(texts_corr_order_index_int)):
|
||||
if i!=j:
|
||||
if regions_ar_less_than_early_min:
|
||||
if regions_ar_less_than_early_min[i]==1:
|
||||
input_multi_visual_modal = np.zeros((input_height,input_width,3)).astype(np.int8)
|
||||
final_f_name = f_name+'_'+str(indexer+indexer_start)
|
||||
order_class_condition = texts_corr_order_index_int[i]-texts_corr_order_index_int[j]
|
||||
if order_class_condition<0:
|
||||
class_type = 1
|
||||
else:
|
||||
class_type = 0
|
||||
|
||||
input_multi_visual_modal[:,:,0] = labels_con[:,:,i]
|
||||
input_multi_visual_modal[:,:,1] = img_poly[:,:,0]
|
||||
input_multi_visual_modal[:,:,2] = labels_con[:,:,j]
|
||||
|
||||
np.save(os.path.join(dir_out_classes,final_f_name+'_missed.npy' ), class_type)
|
||||
|
||||
cv2.imwrite(os.path.join(dir_out_modal_image,final_f_name+'_missed.png' ), input_multi_visual_modal)
|
||||
indexer = indexer+1
|
||||
|
||||
else:
|
||||
input_multi_visual_modal = np.zeros((input_height,input_width,3)).astype(np.int8)
|
||||
final_f_name = f_name+'_'+str(indexer+indexer_start)
|
||||
order_class_condition = texts_corr_order_index_int[i]-texts_corr_order_index_int[j]
|
||||
if order_class_condition<0:
|
||||
class_type = 1
|
||||
else:
|
||||
class_type = 0
|
||||
|
||||
input_multi_visual_modal[:,:,0] = labels_con[:,:,i]
|
||||
input_multi_visual_modal[:,:,1] = img_poly[:,:,0]
|
||||
input_multi_visual_modal[:,:,2] = labels_con[:,:,j]
|
||||
|
||||
np.save(os.path.join(dir_out_classes,final_f_name+'.npy' ), class_type)
|
||||
|
||||
cv2.imwrite(os.path.join(dir_out_modal_image,final_f_name+'.png' ), input_multi_visual_modal)
|
||||
indexer = indexer+1
|
||||
|
||||
|
||||
@main.command()
|
||||
@click.option(
|
||||
"--xml_file",
|
||||
"-xml",
|
||||
help="xml filename",
|
||||
type=click.Path(exists=True, dir_okay=False),
|
||||
)
|
||||
@click.option(
|
||||
"--dir_xml",
|
||||
"-dx",
|
||||
help="directory of GT page-xml files",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
)
|
||||
|
||||
@click.option(
|
||||
"--dir_out",
|
||||
"-o",
|
||||
help="directory where plots will be written",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
)
|
||||
|
||||
@click.option(
|
||||
"--dir_imgs",
|
||||
"-di",
|
||||
help="directory where the overlayed plots will be written", )
|
||||
|
||||
def visualize_reading_order(xml_file, dir_xml, dir_out, dir_imgs):
|
||||
assert xml_file or dir_xml, "A single xml file -xml or a dir of xml files -dx is required not both of them"
|
||||
|
||||
if dir_xml:
|
||||
xml_files_ind = os.listdir(dir_xml)
|
||||
xml_files_ind = [ind_xml for ind_xml in xml_files_ind if ind_xml.endswith('.xml')]
|
||||
else:
|
||||
xml_files_ind = [xml_file]
|
||||
|
||||
indexer_start= 0#55166
|
||||
#min_area = 0.0001
|
||||
|
||||
for ind_xml in tqdm(xml_files_ind):
|
||||
indexer = 0
|
||||
#print(ind_xml)
|
||||
#print('########################')
|
||||
#xml_file = os.path.join(dir_xml,ind_xml )
|
||||
|
||||
if dir_xml:
|
||||
xml_file = os.path.join(dir_xml,ind_xml )
|
||||
f_name = Path(ind_xml).stem
|
||||
else:
|
||||
xml_file = os.path.join(ind_xml )
|
||||
f_name = Path(ind_xml).stem
|
||||
print(f_name, 'f_name')
|
||||
|
||||
#f_name = ind_xml.split('.')[0]
|
||||
_, _, _, file_name, id_paragraph, id_header,co_text_paragraph,co_text_header,tot_region_ref,x_len, y_len,index_tot_regions,img_poly = read_xml(xml_file)
|
||||
|
||||
id_all_text = id_paragraph + id_header
|
||||
co_text_all = co_text_paragraph + co_text_header
|
||||
|
||||
|
||||
cx_main, cy_main, x_min_main, x_max_main, y_min_main, y_max_main, _ = find_new_features_of_contours(co_text_all)
|
||||
|
||||
texts_corr_order_index = [int(index_tot_regions[tot_region_ref.index(i)]) for i in id_all_text ]
|
||||
#texts_corr_order_index_int = [int(x) for x in texts_corr_order_index]
|
||||
|
||||
|
||||
#cx_ordered = np.array(cx_main)[np.array(texts_corr_order_index)]
|
||||
#cx_ordered = cx_ordered.astype(np.int32)
|
||||
|
||||
cx_ordered = [int(val) for (_, val) in sorted(zip(texts_corr_order_index, cx_main), key=lambda x: \
|
||||
x[0], reverse=False)]
|
||||
#cx_ordered = cx_ordered.astype(np.int32)
|
||||
|
||||
cy_ordered = [int(val) for (_, val) in sorted(zip(texts_corr_order_index, cy_main), key=lambda x: \
|
||||
x[0], reverse=False)]
|
||||
#cy_ordered = cy_ordered.astype(np.int32)
|
||||
|
||||
|
||||
color = (0, 0, 255)
|
||||
thickness = 20
|
||||
if dir_imgs:
|
||||
layout = np.zeros( (y_len,x_len,3) )
|
||||
layout = cv2.fillPoly(layout, pts =co_text_all, color=(1,1,1))
|
||||
|
||||
img_file_name_with_format = find_format_of_given_filename_in_dir(dir_imgs, f_name)
|
||||
img = cv2.imread(os.path.join(dir_imgs, img_file_name_with_format))
|
||||
|
||||
overlayed = overlay_layout_on_image(layout, img, cx_ordered, cy_ordered, color, thickness)
|
||||
cv2.imwrite(os.path.join(dir_out, f_name+'.png'), overlayed)
|
||||
|
||||
else:
|
||||
img = np.zeros( (y_len,x_len,3) )
|
||||
img = cv2.fillPoly(img, pts =co_text_all, color=(255,0,0))
|
||||
for i in range(len(cx_ordered)-1):
|
||||
start_point = (int(cx_ordered[i]), int(cy_ordered[i]))
|
||||
end_point = (int(cx_ordered[i+1]), int(cy_ordered[i+1]))
|
||||
img = cv2.arrowedLine(img, start_point, end_point,
|
||||
color, thickness, tipLength = 0.03)
|
||||
|
||||
cv2.imwrite(os.path.join(dir_out, f_name+'.png'), img)
|
||||
|
||||
|
||||
@main.command()
|
||||
@click.option(
|
||||
"--xml_file",
|
||||
"-xml",
|
||||
help="xml filename",
|
||||
type=click.Path(exists=True, dir_okay=False),
|
||||
)
|
||||
@click.option(
|
||||
"--dir_xml",
|
||||
"-dx",
|
||||
help="directory of GT page-xml files",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
)
|
||||
|
||||
@click.option(
|
||||
"--dir_out",
|
||||
"-o",
|
||||
help="directory where plots will be written",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
)
|
||||
|
||||
@click.option(
|
||||
"--dir_imgs",
|
||||
"-di",
|
||||
help="directory of images where textline segmentation will be overlayed", )
|
||||
|
||||
def visualize_textline_segmentation(xml_file, dir_xml, dir_out, dir_imgs):
|
||||
assert xml_file or dir_xml, "A single xml file -xml or a dir of xml files -dx is required not both of them"
|
||||
if dir_xml:
|
||||
xml_files_ind = os.listdir(dir_xml)
|
||||
xml_files_ind = [ind_xml for ind_xml in xml_files_ind if ind_xml.endswith('.xml')]
|
||||
else:
|
||||
xml_files_ind = [xml_file]
|
||||
|
||||
for ind_xml in tqdm(xml_files_ind):
|
||||
indexer = 0
|
||||
#print(ind_xml)
|
||||
#print('########################')
|
||||
xml_file = os.path.join(dir_xml,ind_xml )
|
||||
f_name = Path(ind_xml).stem
|
||||
|
||||
img_file_name_with_format = find_format_of_given_filename_in_dir(dir_imgs, f_name)
|
||||
img = cv2.imread(os.path.join(dir_imgs, img_file_name_with_format))
|
||||
|
||||
co_tetxlines, y_len, x_len = get_textline_contours_for_visualization(xml_file)
|
||||
|
||||
added_image = visualize_image_from_contours(co_tetxlines, img)
|
||||
|
||||
cv2.imwrite(os.path.join(dir_out, f_name+'.png'), added_image)
|
||||
|
||||
|
||||
|
||||
@main.command()
|
||||
@click.option(
|
||||
"--xml_file",
|
||||
"-xml",
|
||||
help="xml filename",
|
||||
type=click.Path(exists=True, dir_okay=False),
|
||||
)
|
||||
@click.option(
|
||||
"--dir_xml",
|
||||
"-dx",
|
||||
help="directory of GT page-xml files",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
)
|
||||
|
||||
@click.option(
|
||||
"--dir_out",
|
||||
"-o",
|
||||
help="directory where plots will be written",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
)
|
||||
|
||||
@click.option(
|
||||
"--dir_imgs",
|
||||
"-di",
|
||||
help="directory of images where textline segmentation will be overlayed", )
|
||||
|
||||
def visualize_layout_segmentation(xml_file, dir_xml, dir_out, dir_imgs):
|
||||
assert xml_file or dir_xml, "A single xml file -xml or a dir of xml files -dx is required not both of them"
|
||||
if dir_xml:
|
||||
xml_files_ind = os.listdir(dir_xml)
|
||||
xml_files_ind = [ind_xml for ind_xml in xml_files_ind if ind_xml.endswith('.xml')]
|
||||
else:
|
||||
xml_files_ind = [xml_file]
|
||||
|
||||
for ind_xml in tqdm(xml_files_ind):
|
||||
indexer = 0
|
||||
#print(ind_xml)
|
||||
#print('########################')
|
||||
if dir_xml:
|
||||
xml_file = os.path.join(dir_xml,ind_xml )
|
||||
f_name = Path(ind_xml).stem
|
||||
else:
|
||||
xml_file = os.path.join(ind_xml )
|
||||
f_name = Path(ind_xml).stem
|
||||
print(f_name, 'f_name')
|
||||
|
||||
img_file_name_with_format = find_format_of_given_filename_in_dir(dir_imgs, f_name)
|
||||
img = cv2.imread(os.path.join(dir_imgs, img_file_name_with_format))
|
||||
|
||||
co_text, co_graphic, co_sep, co_img, co_table, co_map, co_noise, y_len, x_len = get_layout_contours_for_visualization(xml_file)
|
||||
|
||||
|
||||
added_image = visualize_image_from_contours_layout(co_text['paragraph'], co_text['header']+co_text['heading'], co_text['drop-capital'], co_sep, co_img, co_text['marginalia'], co_table, img)
|
||||
|
||||
cv2.imwrite(os.path.join(dir_out, f_name+'.png'), added_image)
|
||||
|
||||
|
||||
|
||||
|
||||
@main.command()
|
||||
@click.option(
|
||||
"--xml_file",
|
||||
"-xml",
|
||||
help="xml filename",
|
||||
type=click.Path(exists=True, dir_okay=False),
|
||||
)
|
||||
@click.option(
|
||||
"--dir_xml",
|
||||
"-dx",
|
||||
help="directory of GT page-xml files",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
)
|
||||
|
||||
@click.option(
|
||||
"--dir_out",
|
||||
"-o",
|
||||
help="directory where plots will be written",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
)
|
||||
|
||||
|
||||
def visualize_ocr_text(xml_file, dir_xml, dir_out):
|
||||
assert xml_file or dir_xml, "A single xml file -xml or a dir of xml files -dx is required not both of them"
|
||||
if dir_xml:
|
||||
xml_files_ind = os.listdir(dir_xml)
|
||||
xml_files_ind = [ind_xml for ind_xml in xml_files_ind if ind_xml.endswith('.xml')]
|
||||
else:
|
||||
xml_files_ind = [xml_file]
|
||||
|
||||
font_path = "Charis-7.000/Charis-Regular.ttf" # Make sure this file exists!
|
||||
font = ImageFont.truetype(font_path, 40)
|
||||
|
||||
for ind_xml in tqdm(xml_files_ind):
|
||||
indexer = 0
|
||||
#print(ind_xml)
|
||||
#print('########################')
|
||||
if dir_xml:
|
||||
xml_file = os.path.join(dir_xml,ind_xml )
|
||||
f_name = Path(ind_xml).stem
|
||||
else:
|
||||
xml_file = os.path.join(ind_xml )
|
||||
f_name = Path(ind_xml).stem
|
||||
print(f_name, 'f_name')
|
||||
|
||||
co_tetxlines, y_len, x_len, ocr_texts = get_textline_contours_and_ocr_text(xml_file)
|
||||
|
||||
total_bb_coordinates = []
|
||||
|
||||
image_text = Image.new("RGB", (x_len, y_len), "white")
|
||||
draw = ImageDraw.Draw(image_text)
|
||||
|
||||
|
||||
|
||||
for index, cnt in enumerate(co_tetxlines):
|
||||
x,y,w,h = cv2.boundingRect(cnt)
|
||||
#total_bb_coordinates.append([x,y,w,h])
|
||||
|
||||
#fit_text_single_line
|
||||
|
||||
#x_bb = bb_ind[0]
|
||||
#y_bb = bb_ind[1]
|
||||
#w_bb = bb_ind[2]
|
||||
#h_bb = bb_ind[3]
|
||||
if ocr_texts[index]:
|
||||
|
||||
|
||||
is_vertical = h > 2*w # Check orientation
|
||||
font = fit_text_single_line(draw, ocr_texts[index], font_path, w, int(h*0.4) )
|
||||
|
||||
if is_vertical:
|
||||
|
||||
vertical_font = fit_text_single_line(draw, ocr_texts[index], font_path, h, int(w * 0.8))
|
||||
|
||||
text_img = Image.new("RGBA", (h, w), (255, 255, 255, 0)) # Note: dimensions are swapped
|
||||
text_draw = ImageDraw.Draw(text_img)
|
||||
text_draw.text((0, 0), ocr_texts[index], font=vertical_font, fill="black")
|
||||
|
||||
# Rotate text image by 90 degrees
|
||||
rotated_text = text_img.rotate(90, expand=1)
|
||||
|
||||
# Calculate paste position (centered in bbox)
|
||||
paste_x = x + (w - rotated_text.width) // 2
|
||||
paste_y = y + (h - rotated_text.height) // 2
|
||||
|
||||
image_text.paste(rotated_text, (paste_x, paste_y), rotated_text) # Use rotated image as mask
|
||||
else:
|
||||
text_bbox = draw.textbbox((0, 0), ocr_texts[index], font=font)
|
||||
text_width = text_bbox[2] - text_bbox[0]
|
||||
text_height = text_bbox[3] - text_bbox[1]
|
||||
|
||||
text_x = x + (w - text_width) // 2 # Center horizontally
|
||||
text_y = y + (h - text_height) // 2 # Center vertically
|
||||
|
||||
# Draw the text
|
||||
draw.text((text_x, text_y), ocr_texts[index], fill="black", font=font)
|
||||
image_text.save(os.path.join(dir_out, f_name+'.png'))
|
||||
File diff suppressed because it is too large
Load diff
|
|
@ -1,682 +0,0 @@
|
|||
"""
|
||||
Tool to load model and predict for given image.
|
||||
"""
|
||||
|
||||
import sys
|
||||
import os
|
||||
from typing import Tuple
|
||||
import warnings
|
||||
import json
|
||||
|
||||
import click
|
||||
import numpy as np
|
||||
from numpy._typing import NDArray
|
||||
import cv2
|
||||
import xml.etree.ElementTree as ET
|
||||
|
||||
os.environ['TF_USE_LEGACY_KERAS'] = '1' # avoid Keras 3 after TF 2.15
|
||||
import tensorflow as tf
|
||||
from tensorflow.keras.models import Model, load_model
|
||||
from tensorflow.keras.layers import StringLookup
|
||||
|
||||
from .gt_gen_utils import (
|
||||
filter_contours_area_of_image,
|
||||
find_new_features_of_contours,
|
||||
read_xml,
|
||||
resize_image,
|
||||
update_list_and_return_first_with_length_bigger_than_one
|
||||
)
|
||||
from ..patch_encoder import (
|
||||
PatchEncoder,
|
||||
Patches
|
||||
)
|
||||
from .metrics import (
|
||||
soft_dice_loss,
|
||||
weighted_categorical_crossentropy,
|
||||
)
|
||||
from.utils import scale_padd_image_for_ocr
|
||||
from ..utils.utils_ocr import decode_batch_predictions
|
||||
|
||||
|
||||
with warnings.catch_warnings():
|
||||
warnings.simplefilter("ignore")
|
||||
|
||||
class SBBPredict:
|
||||
def __init__(self,
|
||||
image,
|
||||
dir_in,
|
||||
model,
|
||||
task,
|
||||
config_params_model,
|
||||
patches,
|
||||
save,
|
||||
save_layout,
|
||||
ground_truth,
|
||||
xml_file,
|
||||
cpu,
|
||||
out,
|
||||
min_area,
|
||||
):
|
||||
self.image=image
|
||||
self.dir_in=dir_in
|
||||
self.patches=patches
|
||||
self.save=save
|
||||
self.save_layout=save_layout
|
||||
self.model_dir=model
|
||||
self.ground_truth=ground_truth
|
||||
self.task=task
|
||||
self.config_params_model=config_params_model
|
||||
self.xml_file = xml_file
|
||||
self.out = out
|
||||
self.cpu = cpu
|
||||
if min_area:
|
||||
self.min_area = float(min_area)
|
||||
else:
|
||||
self.min_area = 0
|
||||
|
||||
def resize_image(self,img_in,input_height,input_width):
|
||||
return cv2.resize(img_in, (input_width,
|
||||
input_height),
|
||||
interpolation=cv2.INTER_NEAREST)
|
||||
|
||||
def color_images(self,seg):
|
||||
ann_u=range(self.n_classes)
|
||||
if len(np.shape(seg))==3:
|
||||
seg=seg[:,:,0]
|
||||
|
||||
seg_img=np.zeros((np.shape(seg)[0],np.shape(seg)[1],3)).astype(np.uint8)
|
||||
|
||||
for c in ann_u:
|
||||
c=int(c)
|
||||
seg_img[:,:,0][seg==c]=c
|
||||
seg_img[:,:,1][seg==c]=c
|
||||
seg_img[:,:,2][seg==c]=c
|
||||
return seg_img
|
||||
|
||||
def IoU(self,Yi,y_predi):
|
||||
## mean Intersection over Union
|
||||
## Mean IoU = TP/(FN + TP + FP)
|
||||
|
||||
IoUs = []
|
||||
Nclass = np.unique(Yi)
|
||||
for c in Nclass:
|
||||
TP = np.sum( (Yi == c)&(y_predi==c) )
|
||||
FP = np.sum( (Yi != c)&(y_predi==c) )
|
||||
FN = np.sum( (Yi == c)&(y_predi != c))
|
||||
IoU = TP/float(TP + FP + FN)
|
||||
if self.n_classes>2:
|
||||
print("class {:02.0f}: #TP={:6.0f}, #FP={:6.0f}, #FN={:5.0f}, IoU={:4.3f}".format(c,TP,FP,FN,IoU))
|
||||
IoUs.append(IoU)
|
||||
if self.n_classes>2:
|
||||
mIoU = np.mean(IoUs)
|
||||
print("_________________")
|
||||
print("Mean IoU: {:4.3f}".format(mIoU))
|
||||
return mIoU
|
||||
elif self.n_classes==2:
|
||||
mIoU = IoUs[1]
|
||||
print("_________________")
|
||||
print("IoU: {:4.3f}".format(mIoU))
|
||||
return mIoU
|
||||
|
||||
def start_new_session_and_model(self):
|
||||
if self.cpu:
|
||||
tf.config.set_visible_devices([], 'GPU')
|
||||
else:
|
||||
try:
|
||||
for device in tf.config.list_physical_devices('GPU'):
|
||||
tf.config.experimental.set_memory_growth(device, True)
|
||||
except:
|
||||
print("no GPU device available", file=sys.stderr)
|
||||
|
||||
if self.task == "cnn-rnn-ocr":
|
||||
self.model = Model(
|
||||
self.model.get_layer(name = "image").input,
|
||||
self.model.get_layer(name = "dense2").output)
|
||||
else:
|
||||
self.model = load_model(self.model_dir, compile=False,
|
||||
custom_objects={"PatchEncoder": PatchEncoder,
|
||||
"Patches": Patches})
|
||||
|
||||
##if self.weights_dir!=None:
|
||||
##self.model.load_weights(self.weights_dir)
|
||||
|
||||
assert isinstance(self.model, Model)
|
||||
if self.task != 'classification' and self.task != 'reading_order':
|
||||
last = self.model.layers[-1]
|
||||
self.img_height = last.output_shape[1]
|
||||
self.img_width = last.output_shape[2]
|
||||
self.n_classes = last.output_shape[3]
|
||||
|
||||
def visualize_model_output(self, prediction, img, task) -> Tuple[NDArray, NDArray]:
|
||||
if task == "binarization":
|
||||
prediction = prediction * -1
|
||||
prediction = prediction + 1
|
||||
added_image = prediction * 255
|
||||
layout_only = None
|
||||
else:
|
||||
unique_classes = np.unique(prediction[:,:,0])
|
||||
rgb_colors = {'0' : [255, 255, 255],
|
||||
'1' : [255, 0, 0],
|
||||
'2' : [255, 125, 0],
|
||||
'3' : [255, 0, 125],
|
||||
'4' : [125, 125, 125],
|
||||
'5' : [125, 125, 0],
|
||||
'6' : [0, 125, 255],
|
||||
'7' : [0, 125, 0],
|
||||
'8' : [125, 125, 125],
|
||||
'9' : [0, 125, 255],
|
||||
'10' : [125, 0, 125],
|
||||
'11' : [0, 255, 0],
|
||||
'12' : [0, 0, 255],
|
||||
'13' : [0, 255, 255],
|
||||
'14' : [255, 125, 125],
|
||||
'15' : [255, 0, 255]}
|
||||
|
||||
layout_only = np.zeros(prediction.shape)
|
||||
for unq_class in unique_classes:
|
||||
where = prediction[:,:,0]==unq_class
|
||||
rgb_class_unique = rgb_colors[str(int(unq_class))]
|
||||
layout_only[:,:,0][where] = rgb_class_unique[0]
|
||||
layout_only[:,:,1][where] = rgb_class_unique[1]
|
||||
layout_only[:,:,2][where] = rgb_class_unique[2]
|
||||
layout_only = layout_only.astype(np.int32)
|
||||
|
||||
img = self.resize_image(img, layout_only.shape[0], layout_only.shape[1])
|
||||
img = img.astype(np.int32)
|
||||
|
||||
added_image = cv2.addWeighted(img,0.5,layout_only,0.1,0)
|
||||
|
||||
assert isinstance(added_image, np.ndarray)
|
||||
assert isinstance(layout_only, np.ndarray)
|
||||
return added_image, layout_only
|
||||
|
||||
def predict(self, image_dir):
|
||||
assert isinstance(self.model, Model)
|
||||
if self.task == 'classification':
|
||||
classes_names = self.config_params_model['classification_classes_name']
|
||||
img_1ch = cv2.imread(image_dir, 0) / 255.0
|
||||
img_1ch = cv2.resize(img_1ch, (self.config_params_model['input_height'],
|
||||
self.config_params_model['input_width']),
|
||||
interpolation=cv2.INTER_NEAREST)
|
||||
img_in = np.zeros((1, img_1ch.shape[0], img_1ch.shape[1], 3))
|
||||
img_in[0, :, :, 0] = img_1ch[:, :]
|
||||
img_in[0, :, :, 1] = img_1ch[:, :]
|
||||
img_in[0, :, :, 2] = img_1ch[:, :]
|
||||
|
||||
label_p_pred = self.model.predict(img_in, verbose='0')
|
||||
index_class = np.argmax(label_p_pred[0])
|
||||
|
||||
print("Predicted Class: {}".format(classes_names[str(int(index_class))]))
|
||||
|
||||
elif self.task == "cnn-rnn-ocr":
|
||||
img=cv2.imread(image_dir)
|
||||
img = scale_padd_image_for_ocr(img, self.config_params_model['input_height'], self.config_params_model['input_width'])
|
||||
|
||||
img = img / 255.
|
||||
|
||||
with open(os.path.join(self.model_dir, "characters_org.txt"), 'r') as char_txt_f:
|
||||
characters = json.load(char_txt_f)
|
||||
|
||||
AUTOTUNE = tf.data.AUTOTUNE
|
||||
|
||||
# Mapping characters to integers.
|
||||
char_to_num = StringLookup(vocabulary=list(characters), mask_token=None)
|
||||
|
||||
# Mapping integers back to original characters.
|
||||
num_to_char = StringLookup(
|
||||
vocabulary=char_to_num.get_vocabulary(), mask_token=None, invert=True
|
||||
)
|
||||
preds = self.model.predict(img.reshape(1, img.shape[0], img.shape[1], img.shape[2]), verbose=0)
|
||||
pred_texts = decode_batch_predictions(preds, num_to_char)
|
||||
pred_texts = pred_texts[0].replace("[UNK]", "")
|
||||
return pred_texts
|
||||
|
||||
|
||||
elif self.task == 'reading_order':
|
||||
img_height = self.config_params_model['input_height']
|
||||
img_width = self.config_params_model['input_width']
|
||||
|
||||
tree_xml, root_xml, ps_bbox, file_name, \
|
||||
id_paragraph, id_header, \
|
||||
co_text_paragraph, co_text_header, \
|
||||
tot_region_ref, x_len, y_len, index_tot_regions, \
|
||||
img_poly = read_xml(self.xml_file)
|
||||
_, cy_main, x_min_main, x_max_main, y_min_main, y_max_main, _ = \
|
||||
find_new_features_of_contours(co_text_header)
|
||||
|
||||
img_header_and_sep = np.zeros((y_len,x_len), dtype='uint8')
|
||||
for j in range(len(cy_main)):
|
||||
img_header_and_sep[int(y_max_main[j]): int(y_max_main[j]) + 12,
|
||||
int(x_min_main[j]): int(x_max_main[j])] = 1
|
||||
|
||||
co_text_all = co_text_paragraph + co_text_header
|
||||
id_all_text = id_paragraph + id_header
|
||||
|
||||
##texts_corr_order_index = [index_tot_regions[tot_region_ref.index(i)] for i in id_all_text ]
|
||||
##texts_corr_order_index_int = [int(x) for x in texts_corr_order_index]
|
||||
texts_corr_order_index_int = list(np.array(range(len(co_text_all))))
|
||||
|
||||
#print(texts_corr_order_index_int)
|
||||
|
||||
max_area = 1
|
||||
#print(np.shape(co_text_all[0]), len( np.shape(co_text_all[0]) ),'co_text_all')
|
||||
#co_text_all = filter_contours_area_of_image_tables(img_poly, co_text_all, _, max_area, min_area)
|
||||
#print(co_text_all,'co_text_all')
|
||||
co_text_all, texts_corr_order_index_int, _ = filter_contours_area_of_image(
|
||||
img_poly, co_text_all, texts_corr_order_index_int, max_area, self.min_area)
|
||||
|
||||
#print(texts_corr_order_index_int)
|
||||
|
||||
#co_text_all = [co_text_all[index] for index in texts_corr_order_index_int]
|
||||
id_all_text = [id_all_text[index] for index in texts_corr_order_index_int]
|
||||
|
||||
labels_con = np.zeros((y_len,x_len,len(co_text_all)),dtype='uint8')
|
||||
for i in range(len(co_text_all)):
|
||||
img_label = np.zeros((y_len,x_len,3),dtype='uint8')
|
||||
img_label=cv2.fillPoly(img_label, pts =[co_text_all[i]], color=(1,1,1))
|
||||
labels_con[:,:,i] = img_label[:,:,0]
|
||||
|
||||
if ps_bbox:
|
||||
labels_con = labels_con[ps_bbox[1]:ps_bbox[3],
|
||||
ps_bbox[0]:ps_bbox[2], :]
|
||||
img_poly = img_poly[ps_bbox[1]:ps_bbox[3],
|
||||
ps_bbox[0]:ps_bbox[2], :]
|
||||
img_header_and_sep = img_header_and_sep[ps_bbox[1]:ps_bbox[3],
|
||||
ps_bbox[0]:ps_bbox[2]]
|
||||
|
||||
|
||||
|
||||
img3= np.copy(img_poly)
|
||||
labels_con = resize_image(labels_con, img_height, img_width)
|
||||
|
||||
img_header_and_sep = resize_image(img_header_and_sep, img_height, img_width)
|
||||
|
||||
img3= resize_image (img3, img_height, img_width)
|
||||
img3 = img3.astype(np.uint16)
|
||||
|
||||
inference_bs = 1#4
|
||||
|
||||
input_1= np.zeros( (inference_bs, img_height, img_width,3))
|
||||
|
||||
|
||||
starting_list_of_regions = [list(range(labels_con.shape[2]))]
|
||||
|
||||
index_update = 0
|
||||
index_selected = starting_list_of_regions[0]
|
||||
|
||||
scalibility_num = 0
|
||||
while index_update>=0:
|
||||
ij_list = starting_list_of_regions[index_update]
|
||||
i = ij_list[0]
|
||||
ij_list.pop(0)
|
||||
|
||||
|
||||
pr_list = []
|
||||
post_list = []
|
||||
|
||||
batch_counter = 0
|
||||
tot_counter = 1
|
||||
|
||||
tot_iteration = len(ij_list)
|
||||
full_bs_ite= tot_iteration//inference_bs
|
||||
last_bs = tot_iteration % inference_bs
|
||||
|
||||
jbatch_indexer =[]
|
||||
for j in ij_list:
|
||||
img1= np.repeat(labels_con[:,:,i][:, :, np.newaxis], 3, axis=2)
|
||||
img2 = np.repeat(labels_con[:,:,j][:, :, np.newaxis], 3, axis=2)
|
||||
|
||||
|
||||
img2[:,:,0][img3[:,:,0]==5] = 2
|
||||
img2[:,:,0][img_header_and_sep[:,:]==1] = 3
|
||||
|
||||
|
||||
|
||||
img1[:,:,0][img3[:,:,0]==5] = 2
|
||||
img1[:,:,0][img_header_and_sep[:,:]==1] = 3
|
||||
|
||||
#input_1= np.zeros( (height1, width1,3))
|
||||
|
||||
|
||||
jbatch_indexer.append(j)
|
||||
|
||||
input_1[batch_counter,:,:,0] = img1[:,:,0]/3.
|
||||
input_1[batch_counter,:,:,2] = img2[:,:,0]/3.
|
||||
input_1[batch_counter,:,:,1] = img3[:,:,0]/5.
|
||||
#input_1[batch_counter,:,:,:]= np.zeros( (batch_counter, height1, width1,3))
|
||||
batch_counter = batch_counter+1
|
||||
|
||||
#input_1[:,:,0] = img1[:,:,0]/3.
|
||||
#input_1[:,:,2] = img2[:,:,0]/3.
|
||||
#input_1[:,:,1] = img3[:,:,0]/5.
|
||||
|
||||
if batch_counter==inference_bs or ( (tot_counter//inference_bs)==full_bs_ite and tot_counter%inference_bs==last_bs):
|
||||
y_pr = self.model.predict(input_1 , verbose='0')
|
||||
scalibility_num = scalibility_num+1
|
||||
|
||||
if batch_counter==inference_bs:
|
||||
iteration_batches = inference_bs
|
||||
else:
|
||||
iteration_batches = last_bs
|
||||
for jb in range(iteration_batches):
|
||||
if y_pr[jb][0]>=0.5:
|
||||
post_list.append(jbatch_indexer[jb])
|
||||
else:
|
||||
pr_list.append(jbatch_indexer[jb])
|
||||
|
||||
batch_counter = 0
|
||||
jbatch_indexer = []
|
||||
|
||||
tot_counter = tot_counter+1
|
||||
|
||||
starting_list_of_regions, index_update = update_list_and_return_first_with_length_bigger_than_one(index_update, i, pr_list, post_list,starting_list_of_regions)
|
||||
|
||||
|
||||
index_sort = [i[0] for i in starting_list_of_regions ]
|
||||
|
||||
id_all_text = np.array(id_all_text)[index_sort]
|
||||
|
||||
alltags=[elem.tag for elem in root_xml.iter()]
|
||||
|
||||
|
||||
|
||||
link=alltags[0].split('}')[0]+'}'
|
||||
name_space = alltags[0].split('}')[0]
|
||||
name_space = name_space.split('{')[1]
|
||||
|
||||
page_element = root_xml.find(link+'Page')
|
||||
assert isinstance(page_element, ET.Element)
|
||||
|
||||
"""
|
||||
ro_subelement = ET.SubElement(page_element, 'ReadingOrder')
|
||||
#print(page_element, 'page_element')
|
||||
|
||||
#new_element = ET.Element('ReadingOrder')
|
||||
|
||||
new_element_element = ET.Element('OrderedGroup')
|
||||
new_element_element.set('id', "ro357564684568544579089")
|
||||
|
||||
for index, id_text in enumerate(id_all_text):
|
||||
new_element_2 = ET.Element('RegionRefIndexed')
|
||||
new_element_2.set('regionRef', id_all_text[index])
|
||||
new_element_2.set('index', str(index_sort[index]))
|
||||
|
||||
new_element_element.append(new_element_2)
|
||||
|
||||
ro_subelement.append(new_element_element)
|
||||
"""
|
||||
##ro_subelement = ET.SubElement(page_element, 'ReadingOrder')
|
||||
|
||||
ro_subelement = ET.Element('ReadingOrder')
|
||||
|
||||
ro_subelement2 = ET.SubElement(ro_subelement, 'OrderedGroup')
|
||||
ro_subelement2.set('id', "ro357564684568544579089")
|
||||
|
||||
for index, id_text in enumerate(id_all_text):
|
||||
new_element_2 = ET.SubElement(ro_subelement2, 'RegionRefIndexed')
|
||||
new_element_2.set('regionRef', id_all_text[index])
|
||||
new_element_2.set('index', str(index))
|
||||
|
||||
if (link+'PrintSpace' in alltags) or (link+'Border' in alltags):
|
||||
page_element.insert(1, ro_subelement)
|
||||
else:
|
||||
page_element.insert(0, ro_subelement)
|
||||
|
||||
alltags=[elem.tag for elem in root_xml.iter()]
|
||||
|
||||
ET.register_namespace("",name_space)
|
||||
tree_xml.write(os.path.join(self.out, file_name+'.xml'),xml_declaration=True,method='xml',encoding="utf8",default_namespace=None)
|
||||
#tree_xml.write('library2.xml')
|
||||
|
||||
else:
|
||||
if self.patches:
|
||||
#def textline_contours(img,input_width,input_height,n_classes,model):
|
||||
|
||||
img=cv2.imread(image_dir)
|
||||
self.img_org = np.copy(img)
|
||||
|
||||
if img.shape[0] < self.img_height:
|
||||
img = self.resize_image(img, self.img_height, img.shape[1])
|
||||
|
||||
if img.shape[1] < self.img_width:
|
||||
img = self.resize_image(img, img.shape[0], self.img_width)
|
||||
|
||||
margin = int(0.1 * self.img_width)
|
||||
width_mid = self.img_width - 2 * margin
|
||||
height_mid = self.img_height - 2 * margin
|
||||
img = img / float(255.0)
|
||||
|
||||
img_h = img.shape[0]
|
||||
img_w = img.shape[1]
|
||||
|
||||
prediction_true = np.zeros((img_h, img_w, 3))
|
||||
nxf = img_w / float(width_mid)
|
||||
nyf = img_h / float(height_mid)
|
||||
|
||||
nxf = int(nxf) + 1 if nxf > int(nxf) else int(nxf)
|
||||
nyf = int(nyf) + 1 if nyf > int(nyf) else int(nyf)
|
||||
|
||||
for i in range(nxf):
|
||||
for j in range(nyf):
|
||||
if i == 0:
|
||||
index_x_d = i * width_mid
|
||||
index_x_u = index_x_d + self.img_width
|
||||
else:
|
||||
index_x_d = i * width_mid
|
||||
index_x_u = index_x_d + self.img_width
|
||||
if j == 0:
|
||||
index_y_d = j * height_mid
|
||||
index_y_u = index_y_d + self.img_height
|
||||
else:
|
||||
index_y_d = j * height_mid
|
||||
index_y_u = index_y_d + self.img_height
|
||||
|
||||
if index_x_u > img_w:
|
||||
index_x_u = img_w
|
||||
index_x_d = img_w - self.img_width
|
||||
if index_y_u > img_h:
|
||||
index_y_u = img_h
|
||||
index_y_d = img_h - self.img_height
|
||||
|
||||
img_patch = img[index_y_d:index_y_u, index_x_d:index_x_u, :]
|
||||
label_p_pred = self.model.predict(img_patch.reshape(1, img_patch.shape[0], img_patch.shape[1], img_patch.shape[2]),
|
||||
verbose='0')
|
||||
|
||||
if self.task == 'enhancement':
|
||||
seg = label_p_pred[0, :, :, :]
|
||||
seg = seg * 255
|
||||
elif self.task == 'segmentation' or self.task == 'binarization':
|
||||
seg = np.argmax(label_p_pred, axis=3)[0]
|
||||
seg = np.repeat(seg[:, :, np.newaxis], 3, axis=2)
|
||||
else:
|
||||
raise ValueError(f"Unhandled task {self.task}")
|
||||
|
||||
|
||||
if i == 0 and j == 0:
|
||||
seg = seg[0 : seg.shape[0] - margin, 0 : seg.shape[1] - margin]
|
||||
prediction_true[index_y_d + 0 : index_y_u - margin, index_x_d + 0 : index_x_u - margin, :] = seg
|
||||
elif i == nxf - 1 and j == nyf - 1:
|
||||
seg = seg[margin : seg.shape[0] - 0, margin : seg.shape[1] - 0]
|
||||
prediction_true[index_y_d + margin : index_y_u - 0, index_x_d + margin : index_x_u - 0, :] = seg
|
||||
elif i == 0 and j == nyf - 1:
|
||||
seg = seg[margin : seg.shape[0] - 0, 0 : seg.shape[1] - margin]
|
||||
prediction_true[index_y_d + margin : index_y_u - 0, index_x_d + 0 : index_x_u - margin, :] = seg
|
||||
elif i == nxf - 1 and j == 0:
|
||||
seg = seg[0 : seg.shape[0] - margin, margin : seg.shape[1] - 0]
|
||||
prediction_true[index_y_d + 0 : index_y_u - margin, index_x_d + margin : index_x_u - 0, :] = seg
|
||||
elif i == 0 and j != 0 and j != nyf - 1:
|
||||
seg = seg[margin : seg.shape[0] - margin, 0 : seg.shape[1] - margin]
|
||||
prediction_true[index_y_d + margin : index_y_u - margin, index_x_d + 0 : index_x_u - margin, :] = seg
|
||||
elif i == nxf - 1 and j != 0 and j != nyf - 1:
|
||||
seg = seg[margin : seg.shape[0] - margin, margin : seg.shape[1] - 0]
|
||||
prediction_true[index_y_d + margin : index_y_u - margin, index_x_d + margin : index_x_u - 0, :] = seg
|
||||
elif i != 0 and i != nxf - 1 and j == 0:
|
||||
seg = seg[0 : seg.shape[0] - margin, margin : seg.shape[1] - margin]
|
||||
prediction_true[index_y_d + 0 : index_y_u - margin, index_x_d + margin : index_x_u - margin, :] = seg
|
||||
elif i != 0 and i != nxf - 1 and j == nyf - 1:
|
||||
seg = seg[margin : seg.shape[0] - 0, margin : seg.shape[1] - margin]
|
||||
prediction_true[index_y_d + margin : index_y_u - 0, index_x_d + margin : index_x_u - margin, :] = seg
|
||||
else:
|
||||
seg = seg[margin : seg.shape[0] - margin, margin : seg.shape[1] - margin]
|
||||
prediction_true[index_y_d + margin : index_y_u - margin, index_x_d + margin : index_x_u - margin, :] = seg
|
||||
prediction_true = prediction_true.astype(int)
|
||||
prediction_true = cv2.resize(prediction_true, (self.img_org.shape[1], self.img_org.shape[0]), interpolation=cv2.INTER_NEAREST)
|
||||
return prediction_true
|
||||
|
||||
else:
|
||||
|
||||
img=cv2.imread(image_dir)
|
||||
self.img_org = np.copy(img)
|
||||
|
||||
width=self.img_width
|
||||
height=self.img_height
|
||||
|
||||
img=img/255.0
|
||||
img=self.resize_image(img,self.img_height,self.img_width)
|
||||
|
||||
|
||||
label_p_pred=self.model.predict(
|
||||
img.reshape(1,img.shape[0],img.shape[1],img.shape[2]))
|
||||
|
||||
if self.task == 'enhancement':
|
||||
seg = label_p_pred[0, :, :, :]
|
||||
seg = seg * 255
|
||||
elif self.task == 'segmentation' or self.task == 'binarization':
|
||||
seg = np.argmax(label_p_pred, axis=3)[0]
|
||||
seg = np.repeat(seg[:, :, np.newaxis], 3, axis=2)
|
||||
else:
|
||||
raise ValueError(f"Unhandled task {self.task}")
|
||||
|
||||
prediction_true = seg.astype(int)
|
||||
|
||||
prediction_true = cv2.resize(prediction_true, (self.img_org.shape[1], self.img_org.shape[0]), interpolation=cv2.INTER_NEAREST)
|
||||
return prediction_true
|
||||
|
||||
|
||||
|
||||
def run(self):
|
||||
self.start_new_session_and_model()
|
||||
if self.image:
|
||||
res=self.predict(image_dir = self.image)
|
||||
|
||||
if self.task == 'classification' or self.task == 'reading_order':
|
||||
pass
|
||||
elif self.task == 'enhancement':
|
||||
if self.save:
|
||||
cv2.imwrite(self.save,res)
|
||||
elif self.task == "cnn-rnn-ocr":
|
||||
print(f"Detected text: {res}")
|
||||
else:
|
||||
img_seg_overlayed, only_layout = self.visualize_model_output(res, self.img_org, self.task)
|
||||
if self.save:
|
||||
cv2.imwrite(self.save,img_seg_overlayed)
|
||||
if self.save_layout:
|
||||
cv2.imwrite(self.save_layout, only_layout)
|
||||
|
||||
if self.ground_truth:
|
||||
gt_img=cv2.imread(self.ground_truth)
|
||||
self.IoU(gt_img[:,:,0],res[:,:,0])
|
||||
|
||||
else:
|
||||
ls_images = os.listdir(self.dir_in)
|
||||
for ind_image in ls_images:
|
||||
f_name = ind_image.split('.')[0]
|
||||
image_dir = os.path.join(self.dir_in, ind_image)
|
||||
res=self.predict(image_dir)
|
||||
|
||||
if self.task == 'classification' or self.task == 'reading_order':
|
||||
pass
|
||||
elif self.task == 'enhancement':
|
||||
self.save = os.path.join(self.out, f_name+'.png')
|
||||
cv2.imwrite(self.save,res)
|
||||
elif self.task == "cnn-rnn-ocr":
|
||||
print(f"Detected text for file name {f_name} is: {res}")
|
||||
else:
|
||||
img_seg_overlayed, only_layout = self.visualize_model_output(res, self.img_org, self.task)
|
||||
self.save = os.path.join(self.out, f_name+'_overlayed.png')
|
||||
cv2.imwrite(self.save,img_seg_overlayed)
|
||||
self.save_layout = os.path.join(self.out, f_name+'_layout.png')
|
||||
cv2.imwrite(self.save_layout, only_layout)
|
||||
|
||||
if self.ground_truth:
|
||||
gt_img=cv2.imread(self.ground_truth)
|
||||
self.IoU(gt_img[:,:,0],res[:,:,0])
|
||||
|
||||
|
||||
|
||||
@click.command()
|
||||
@click.option(
|
||||
"--image",
|
||||
"-i",
|
||||
help="image filename",
|
||||
type=click.Path(exists=True, dir_okay=False),
|
||||
)
|
||||
@click.option(
|
||||
"--dir_in",
|
||||
"-di",
|
||||
help="directory of images",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
)
|
||||
@click.option(
|
||||
"--out",
|
||||
"-o",
|
||||
help="output directory where xml with detected reading order will be written.",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
)
|
||||
@click.option(
|
||||
"--patches/--no-patches",
|
||||
"-p/-nop",
|
||||
is_flag=True,
|
||||
help="if this parameter set to true, this tool will try to do inference in patches.",
|
||||
)
|
||||
@click.option(
|
||||
"--save",
|
||||
"-s",
|
||||
help="save prediction as a png file in current folder.",
|
||||
)
|
||||
@click.option(
|
||||
"--save_layout",
|
||||
"-sl",
|
||||
help="save layout prediction only as a png file in current folder.",
|
||||
)
|
||||
@click.option(
|
||||
"--model",
|
||||
"-m",
|
||||
help="directory of models",
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
required=True,
|
||||
)
|
||||
@click.option(
|
||||
"--ground_truth",
|
||||
"-gt",
|
||||
help="ground truth directory if you want to see the iou of prediction.",
|
||||
)
|
||||
@click.option(
|
||||
"--xml_file",
|
||||
"-xml",
|
||||
help="xml file with layout coordinates that reading order detection will be implemented on. The result will be written in the same xml file.",
|
||||
)
|
||||
@click.option(
|
||||
"--cpu",
|
||||
"-cpu",
|
||||
help="For OCR, the default device is the GPU. If this parameter is set to true, inference will be performed on the CPU",
|
||||
is_flag=True,
|
||||
)
|
||||
@click.option(
|
||||
"--min_area",
|
||||
"-min",
|
||||
help="min area size of regions considered for reading order detection. The default value is zero and means that all text regions are considered for reading order.",
|
||||
)
|
||||
def main(image, dir_in, model, patches, save, save_layout, ground_truth, xml_file, cpu, out, min_area):
|
||||
assert image or dir_in, "Either a single image -i or a dir_in -di input is required"
|
||||
with open(os.path.join(model,'config.json')) as f:
|
||||
config_params_model = json.load(f)
|
||||
task = config_params_model['task']
|
||||
if task not in ['classification', 'reading_order', "cnn-rnn-ocr"]:
|
||||
assert not image or save, "For segmentation or binarization, an input single image -i also requires an output filename -s"
|
||||
assert not dir_in or out, "For segmentation or binarization, an input directory -di also requires an output directory -o"
|
||||
x = SBBPredict(image, dir_in, model, task, config_params_model,
|
||||
patches, save, save_layout, ground_truth, xml_file,
|
||||
cpu, out, min_area)
|
||||
x.run()
|
||||
|
||||
|
|
@ -1,526 +0,0 @@
|
|||
import os
|
||||
|
||||
os.environ['TF_USE_LEGACY_KERAS'] = '1' # avoid Keras 3 after TF 2.15
|
||||
import tensorflow as tf
|
||||
from tensorflow.keras import backend as K
|
||||
from tensorflow.keras.metrics import Metric, MeanMetricWrapper, get
|
||||
from tensorflow.keras.initializers import Zeros
|
||||
from tensorflow_addons.image import connected_components
|
||||
import numpy as np
|
||||
|
||||
|
||||
EPS = K.epsilon()
|
||||
|
||||
def focal_loss(gamma=2., alpha=4., epsilon=EPS):
|
||||
gamma = float(gamma)
|
||||
alpha = float(alpha)
|
||||
|
||||
def focal_loss_fixed(y_true, y_pred):
|
||||
"""Focal loss for multi-classification
|
||||
FL(p_t)=-alpha(1-p_t)^{gamma}ln(p_t)
|
||||
Notice: y_pred is probability after softmax
|
||||
gradient is d(Fl)/d(p_t) not d(Fl)/d(x) as described in paper
|
||||
d(Fl)/d(p_t) * [p_t(1-p_t)] = d(Fl)/d(x)
|
||||
Focal Loss for Dense Object Detection
|
||||
https://arxiv.org/abs/1708.02002
|
||||
|
||||
Arguments:
|
||||
y_true {tensor} -- ground truth labels, shape of [batch_size, num_cls]
|
||||
y_pred {tensor} -- model's output, shape of [batch_size, num_cls]
|
||||
|
||||
Keyword Arguments:
|
||||
gamma {float} -- (default: {2.0})
|
||||
alpha {float} -- (default: {4.0})
|
||||
|
||||
Returns:
|
||||
[tensor] -- loss.
|
||||
"""
|
||||
y_true = tf.convert_to_tensor(y_true, tf.float32)
|
||||
y_pred = tf.convert_to_tensor(y_pred, tf.float32)
|
||||
|
||||
model_out = tf.add(y_pred, epsilon)
|
||||
ce = tf.multiply(y_true, -tf.log(model_out))
|
||||
weight = tf.multiply(y_true, tf.pow(tf.subtract(1., model_out), gamma))
|
||||
fl = tf.multiply(alpha, tf.multiply(weight, ce))
|
||||
reduced_fl = tf.reduce_max(fl, axis=1)
|
||||
return tf.reduce_mean(reduced_fl)
|
||||
|
||||
return focal_loss_fixed
|
||||
|
||||
|
||||
def weighted_categorical_crossentropy(weights=None):
|
||||
""" weighted_categorical_crossentropy
|
||||
|
||||
Args:
|
||||
* weights<ktensor|nparray|list>: crossentropy weights
|
||||
Returns:
|
||||
* weighted categorical crossentropy function
|
||||
"""
|
||||
|
||||
def loss(y_true, y_pred):
|
||||
labels_floats = tf.cast(y_true, tf.float32)
|
||||
per_pixel_loss = tf.nn.sigmoid_cross_entropy_with_logits(labels=labels_floats, logits=y_pred)
|
||||
|
||||
if weights is not None:
|
||||
weight_mask = tf.maximum(tf.reduce_max(tf.constant(
|
||||
np.array(weights, dtype=np.float32)[None, None, None])
|
||||
* labels_floats, axis=-1), 1.0)
|
||||
per_pixel_loss = per_pixel_loss * weight_mask[:, :, :, None]
|
||||
return tf.reduce_mean(per_pixel_loss)
|
||||
|
||||
return loss
|
||||
|
||||
|
||||
def image_categorical_cross_entropy(y_true, y_pred, weights=None):
|
||||
"""
|
||||
:param y_true: tensor of shape (batch_size, height, width) representing the ground truth.
|
||||
:param y_pred: tensor of shape (batch_size, height, width) representing the prediction.
|
||||
:return: The mean cross-entropy on softmaxed tensors.
|
||||
"""
|
||||
|
||||
labels_floats = tf.cast(y_true, tf.float32)
|
||||
per_pixel_loss = tf.nn.sigmoid_cross_entropy_with_logits(labels=labels_floats, logits=y_pred)
|
||||
|
||||
if weights is not None:
|
||||
weight_mask = tf.maximum(
|
||||
tf.reduce_max(tf.constant(
|
||||
np.array(weights, dtype=np.float32)[None, None, None])
|
||||
* labels_floats, axis=-1), 1.0)
|
||||
per_pixel_loss = per_pixel_loss * weight_mask[:, :, :, None]
|
||||
|
||||
return tf.reduce_mean(per_pixel_loss)
|
||||
|
||||
|
||||
def class_tversky(y_true, y_pred):
|
||||
smooth = 1.0 # 1.00
|
||||
|
||||
y_true = K.permute_dimensions(y_true, (3, 1, 2, 0))
|
||||
y_pred = K.permute_dimensions(y_pred, (3, 1, 2, 0))
|
||||
|
||||
y_true_pos = K.batch_flatten(y_true)
|
||||
y_pred_pos = K.batch_flatten(y_pred)
|
||||
true_pos = K.sum(y_true_pos * y_pred_pos, 1)
|
||||
false_neg = K.sum(y_true_pos * (1 - y_pred_pos), 1)
|
||||
false_pos = K.sum((1 - y_true_pos) * y_pred_pos, 1)
|
||||
alpha = 0.2 # 0.5
|
||||
beta = 0.8
|
||||
return (true_pos + smooth) / (true_pos + alpha * false_neg + beta * false_pos + smooth)
|
||||
|
||||
|
||||
def focal_tversky_loss(y_true, y_pred):
|
||||
pt_1 = class_tversky(y_true, y_pred)
|
||||
gamma = 1.3 # 4./3.0#1.3#4.0/3.00# 0.75
|
||||
return K.sum(K.pow((1 - pt_1), gamma))
|
||||
|
||||
|
||||
def generalized_dice_coeff2(y_true, y_pred):
|
||||
n_el = 1
|
||||
for dim in y_true.shape:
|
||||
n_el *= int(dim)
|
||||
n_cl = y_true.shape[-1]
|
||||
w = K.zeros(shape=(n_cl,))
|
||||
w = (K.sum(y_true, axis=(0, 1, 2))) / n_el
|
||||
w = 1 / (w ** 2 + 0.000001)
|
||||
numerator = y_true * y_pred
|
||||
numerator = w * K.sum(numerator, (0, 1, 2))
|
||||
numerator = K.sum(numerator)
|
||||
denominator = y_true + y_pred
|
||||
denominator = w * K.sum(denominator, (0, 1, 2))
|
||||
denominator = K.sum(denominator)
|
||||
return 2 * numerator / denominator
|
||||
|
||||
|
||||
def generalized_dice_coeff(y_true, y_pred):
|
||||
axes = tuple(range(1, len(y_pred.shape) - 1))
|
||||
Ncl = y_pred.shape[-1]
|
||||
w = K.zeros(shape=(Ncl,))
|
||||
w = K.sum(y_true, axis=axes)
|
||||
w = 1 / (w ** 2 + 0.000001)
|
||||
# Compute gen dice coef:
|
||||
numerator = y_true * y_pred
|
||||
numerator = w * K.sum(numerator, axes)
|
||||
numerator = K.sum(numerator)
|
||||
|
||||
denominator = y_true + y_pred
|
||||
denominator = w * K.sum(denominator, axes)
|
||||
denominator = K.sum(denominator)
|
||||
|
||||
gen_dice_coef = 2 * numerator / denominator
|
||||
|
||||
return gen_dice_coef
|
||||
|
||||
|
||||
def generalized_dice_loss(y_true, y_pred):
|
||||
return 1 - generalized_dice_coeff2(y_true, y_pred)
|
||||
|
||||
|
||||
# TODO: document where this is from
|
||||
def soft_dice_loss(y_true, y_pred, epsilon=EPS):
|
||||
"""
|
||||
Soft dice loss calculation for arbitrary batch size, number of classes, and number of spatial dimensions.
|
||||
Assumes the `channels_last` format.
|
||||
|
||||
# Arguments
|
||||
y_true: b x X x Y( x Z...) x c One hot encoding of ground truth
|
||||
y_pred: b x X x Y( x Z...) x c Network output, must sum to 1 over c channel (such as after softmax)
|
||||
epsilon: Used for numerical stability to avoid divide by zero errors
|
||||
|
||||
# References
|
||||
V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation
|
||||
https://arxiv.org/abs/1606.04797
|
||||
More details on Dice loss formulation
|
||||
https://mediatum.ub.tum.de/doc/1395260/1395260.pdf (page 72)
|
||||
|
||||
Adapted from https://github.com/Lasagne/Recipes/issues/99#issuecomment-347775022
|
||||
"""
|
||||
|
||||
# skip the batch and class axis for calculating Dice score
|
||||
axes = tuple(range(1, len(y_pred.shape) - 1))
|
||||
|
||||
numerator = 2. * K.sum(y_pred * y_true, axes)
|
||||
|
||||
denominator = K.sum(K.square(y_pred) + K.square(y_true), axes)
|
||||
return 1.00 - K.mean(numerator / (denominator + epsilon)) # average over classes and batch
|
||||
|
||||
|
||||
# TODO: document where this is from
|
||||
def seg_metrics(y_true, y_pred, metric_name, metric_type='standard', drop_last=True, mean_per_class=False,
|
||||
verbose=False):
|
||||
"""
|
||||
Compute mean metrics of two segmentation masks, via Keras.
|
||||
|
||||
IoU(A,B) = |A & B| / (| A U B|)
|
||||
Dice(A,B) = 2*|A & B| / (|A| + |B|)
|
||||
|
||||
Args:
|
||||
y_true: true masks, one-hot encoded.
|
||||
y_pred: predicted masks, either softmax outputs, or one-hot encoded.
|
||||
metric_name: metric to be computed, either 'iou' or 'dice'.
|
||||
metric_type: one of 'standard' (default), 'soft', 'naive'.
|
||||
In the standard version, y_pred is one-hot encoded and the mean
|
||||
is taken only over classes that are present (in y_true or y_pred).
|
||||
The 'soft' version of the metrics are computed without one-hot
|
||||
encoding y_pred.
|
||||
The 'naive' version return mean metrics where absent classes contribute
|
||||
to the class mean as 1.0 (instead of being dropped from the mean).
|
||||
drop_last = True: boolean flag to drop last class (usually reserved
|
||||
for background class in semantic segmentation)
|
||||
mean_per_class = False: return mean along batch axis for each class.
|
||||
verbose = False: print intermediate results such as intersection, union
|
||||
(as number of pixels).
|
||||
Returns:
|
||||
IoU/Dice of y_true and y_pred, as a float, unless mean_per_class == True
|
||||
in which case it returns the per-class metric, averaged over the batch.
|
||||
|
||||
Inputs are B*W*H*N tensors, with
|
||||
B = batch size,
|
||||
W = width,
|
||||
H = height,
|
||||
N = number of classes
|
||||
"""
|
||||
|
||||
flag_soft = (metric_type == 'soft')
|
||||
flag_naive_mean = (metric_type == 'naive')
|
||||
|
||||
# always assume one or more classes
|
||||
num_classes = K.shape(y_true)[-1]
|
||||
|
||||
if not flag_soft:
|
||||
# get one-hot encoded masks from y_pred (true masks should already be one-hot)
|
||||
y_pred = K.one_hot(K.argmax(y_pred), num_classes)
|
||||
y_true = K.one_hot(K.argmax(y_true), num_classes)
|
||||
|
||||
# if already one-hot, could have skipped above command
|
||||
# keras uses float32 instead of float64, would give error down (but numpy arrays or keras.to_categorical gives float64)
|
||||
y_true = K.cast(y_true, 'float32')
|
||||
y_pred = K.cast(y_pred, 'float32')
|
||||
|
||||
# intersection and union shapes are batch_size * n_classes (values = area in pixels)
|
||||
axes = (1, 2) # W,H axes of each image
|
||||
intersection = K.sum(K.abs(y_true * y_pred), axis=axes)
|
||||
mask_sum = K.sum(K.abs(y_true), axis=axes) + K.sum(K.abs(y_pred), axis=axes)
|
||||
union = mask_sum - intersection # or, np.logical_or(y_pred, y_true) for one-hot
|
||||
|
||||
smooth = .001
|
||||
iou = (intersection + smooth) / (union + smooth)
|
||||
dice = 2 * (intersection + smooth) / (mask_sum + smooth)
|
||||
|
||||
metric = {'iou': iou, 'dice': dice}[metric_name]
|
||||
|
||||
# define mask to be 0 when no pixels are present in either y_true or y_pred, 1 otherwise
|
||||
mask = K.cast(K.not_equal(union, 0), 'float32')
|
||||
|
||||
if drop_last:
|
||||
metric = metric[:, :-1]
|
||||
mask = mask[:, :-1]
|
||||
|
||||
if verbose:
|
||||
print('intersection, union')
|
||||
print(K.eval(intersection), K.eval(union))
|
||||
print(K.eval(intersection / union))
|
||||
|
||||
# return mean metrics: remaining axes are (batch, classes)
|
||||
if flag_naive_mean:
|
||||
return K.mean(metric)
|
||||
|
||||
# take mean only over non-absent classes
|
||||
class_count = K.sum(mask, axis=0)
|
||||
non_zero = tf.greater(class_count, 0)
|
||||
non_zero_sum = tf.boolean_mask(K.sum(metric * mask, axis=0), non_zero)
|
||||
non_zero_count = tf.boolean_mask(class_count, non_zero)
|
||||
|
||||
if verbose:
|
||||
print('Counts of inputs with class present, metrics for non-absent classes')
|
||||
print(K.eval(class_count), K.eval(non_zero_sum / non_zero_count))
|
||||
|
||||
return K.mean(non_zero_sum / non_zero_count)
|
||||
|
||||
|
||||
# TODO: document where this is from
|
||||
# TODO: Why a different implementation than IoU from utils?
|
||||
def mean_iou(y_true, y_pred, **kwargs):
|
||||
"""
|
||||
Compute mean Intersection over Union of two segmentation masks, via Keras.
|
||||
|
||||
Calls metrics_k(y_true, y_pred, metric_name='iou'), see there for allowed kwargs.
|
||||
"""
|
||||
return seg_metrics(y_true, y_pred, metric_name='iou', **kwargs)
|
||||
|
||||
|
||||
def Mean_IOU(y_true, y_pred):
|
||||
nb_classes = K.int_shape(y_pred)[-1]
|
||||
iou = []
|
||||
true_pixels = K.argmax(y_true, axis=-1)
|
||||
pred_pixels = K.argmax(y_pred, axis=-1)
|
||||
void_labels = K.equal(K.sum(y_true, axis=-1), 0)
|
||||
for i in range(0, nb_classes): # exclude first label (background) and last label (void)
|
||||
true_labels = K.equal(true_pixels, i) # & ~void_labels
|
||||
pred_labels = K.equal(pred_pixels, i) # & ~void_labels
|
||||
inter = tf.to_int32(true_labels & pred_labels)
|
||||
union = tf.to_int32(true_labels | pred_labels)
|
||||
legal_batches = K.sum(tf.to_int32(true_labels), axis=1) > 0
|
||||
ious = K.sum(inter, axis=1) / K.sum(union, axis=1)
|
||||
iou.append(K.mean(tf.gather(ious, indices=tf.where(legal_batches)))) # returns average IoU of the same objects
|
||||
iou = tf.stack(iou)
|
||||
legal_labels = ~tf.debugging.is_nan(iou)
|
||||
iou = tf.gather(iou, indices=tf.where(legal_labels))
|
||||
return K.mean(iou)
|
||||
|
||||
|
||||
def iou_vahid(y_true, y_pred):
|
||||
nb_classes = tf.shape(y_true)[-1] + tf.to_int32(1)
|
||||
true_pixels = K.argmax(y_true, axis=-1)
|
||||
pred_pixels = K.argmax(y_pred, axis=-1)
|
||||
iou = []
|
||||
|
||||
for i in tf.range(nb_classes):
|
||||
tp = K.sum(tf.to_int32(K.equal(true_pixels, i) & K.equal(pred_pixels, i)))
|
||||
fp = K.sum(tf.to_int32(K.not_equal(true_pixels, i) & K.equal(pred_pixels, i)))
|
||||
fn = K.sum(tf.to_int32(K.equal(true_pixels, i) & K.not_equal(pred_pixels, i)))
|
||||
iouh = tp / (tp + fp + fn)
|
||||
iou.append(iouh)
|
||||
return K.mean(iou)
|
||||
|
||||
|
||||
# TODO: copy from utils?
|
||||
def IoU_metric(Yi, y_predi):
|
||||
# mean Intersection over Union
|
||||
# Mean IoU = TP/(FN + TP + FP)
|
||||
y_predi = np.argmax(y_predi, axis=3)
|
||||
y_testi = np.argmax(Yi, axis=3)
|
||||
IoUs = []
|
||||
Nclass = int(np.max(Yi)) + 1
|
||||
for c in range(Nclass):
|
||||
TP = np.sum((Yi == c) & (y_predi == c))
|
||||
FP = np.sum((Yi != c) & (y_predi == c))
|
||||
FN = np.sum((Yi == c) & (y_predi != c))
|
||||
IoU = TP / float(TP + FP + FN)
|
||||
IoUs.append(IoU)
|
||||
return K.cast(np.mean(IoUs), dtype='float32')
|
||||
|
||||
|
||||
def IoU_metric_keras(y_true, y_pred):
|
||||
# mean Intersection over Union
|
||||
# Mean IoU = TP/(FN + TP + FP)
|
||||
init = tf.global_variables_initializer()
|
||||
sess = tf.Session()
|
||||
sess.run(init)
|
||||
|
||||
return IoU_metric(y_true.eval(session=sess), y_pred.eval(session=sess))
|
||||
|
||||
|
||||
# TODO: unused, remove?
|
||||
def jaccard_distance_loss(y_true, y_pred, smooth=100):
|
||||
"""
|
||||
Jaccard = (|X & Y|)/ (|X|+ |Y| - |X & Y|)
|
||||
= sum(|A*B|)/(sum(|A|)+sum(|B|)-sum(|A*B|))
|
||||
|
||||
The jaccard distance loss is usefull for unbalanced datasets. This has been
|
||||
shifted so it converges on 0 and is smoothed to avoid exploding or disapearing
|
||||
gradient.
|
||||
|
||||
Ref: https://en.wikipedia.org/wiki/Jaccard_index
|
||||
|
||||
@url: https://gist.github.com/wassname/f1452b748efcbeb4cb9b1d059dce6f96
|
||||
@author: wassname
|
||||
"""
|
||||
intersection = K.sum(K.abs(y_true * y_pred), axis=-1)
|
||||
sum_ = K.sum(K.abs(y_true) + K.abs(y_pred), axis=-1)
|
||||
jac = (intersection + smooth) / (sum_ - intersection + smooth)
|
||||
return (1 - jac) * smooth
|
||||
|
||||
|
||||
def metrics_superposition(*metrics, weights=None):
|
||||
"""
|
||||
return a single metric derived by adding all given metrics
|
||||
|
||||
default weights are uniform
|
||||
"""
|
||||
if weights is None:
|
||||
weights = len(metrics) * [tf.constant(1.0)]
|
||||
def mixed(y_true, y_pred):
|
||||
results = []
|
||||
for metric, weight in zip(metrics, weights):
|
||||
results.append(metric(y_true, y_pred) * weight)
|
||||
return tf.reduce_mean(tf.stack(results), 0)
|
||||
mixed.__name__ = '/'.join(m.__name__ for m in metrics)
|
||||
return mixed
|
||||
|
||||
|
||||
class Superposition(MeanMetricWrapper):
|
||||
def __init__(self, metrics, weights=None, dtype=None):
|
||||
self._metrics = metrics
|
||||
self._weights = weights
|
||||
mixed = metrics_superposition(*metrics, weights=weights)
|
||||
super().__init__(mixed, name=mixed.__name__, dtype=dtype)
|
||||
def get_config(self):
|
||||
return dict(metrics=self._metrics,
|
||||
weights=self._weights,
|
||||
**super().get_config())
|
||||
|
||||
class ConfusionMatrix(Metric):
|
||||
def __init__(self, nlabels=None, nrm="all", name="confusion_matrix", dtype=tf.float32):
|
||||
super().__init__(name=name, dtype=dtype)
|
||||
assert nlabels is not None
|
||||
self._nlabels = nlabels
|
||||
self._shape = (self._nlabels, self._nlabels)
|
||||
self._matrix = self.add_weight(name, shape=self._shape,
|
||||
initializer=Zeros)
|
||||
assert nrm in ("all", "true", "pred", "none")
|
||||
self._nrm = nrm
|
||||
|
||||
def update_state(self, y_true, y_pred, sample_weight=None):
|
||||
y_pred = tf.math.argmax(y_pred, axis=-1)
|
||||
y_true = tf.math.argmax(y_true, axis=-1)
|
||||
|
||||
y_pred = tf.reshape(y_pred, shape=(-1,))
|
||||
y_true = tf.reshape(y_true, shape=(-1,))
|
||||
|
||||
y_pred.shape.assert_is_compatible_with(y_true.shape)
|
||||
confusion = tf.math.confusion_matrix(y_true, y_pred, num_classes=self._nlabels, dtype=self._dtype)
|
||||
|
||||
return self._matrix.assign_add(confusion)
|
||||
|
||||
def result(self):
|
||||
"""normalize"""
|
||||
if self._nrm == "all":
|
||||
denom = tf.math.reduce_sum(self._matrix, axis=(0, 1))
|
||||
elif self._nrm == "true":
|
||||
denom = tf.math.reduce_sum(self._matrix, axis=1, keepdims=True)
|
||||
elif self._nrm == "pred":
|
||||
denom = tf.math.reduce_sum(self._matrix, axis=0, keepdims=True)
|
||||
else:
|
||||
denom = tf.constant(1.0)
|
||||
return tf.math.divide_no_nan(self._matrix, denom)
|
||||
|
||||
def reset_state(self):
|
||||
for v in self.variables:
|
||||
v.assign(tf.zeros(shape=self._shape))
|
||||
|
||||
def get_config(self):
|
||||
return dict(nlabels=self._nlabels,
|
||||
**super().get_config())
|
||||
|
||||
def connected_components_loss(artificial=0):
|
||||
"""
|
||||
metric/loss function capturing the separability of segmentation maps
|
||||
|
||||
For both sides (true and predicted, resp.), computes
|
||||
1. the argmax() of class-wise softmax input (i.e. the segmentation map)
|
||||
2. the connected components (i.e. the instance label map)
|
||||
3. the max() (i.e. the highest label = nr of components)
|
||||
|
||||
The original idea was to then calculate a regression formula
|
||||
between those two targets. But it is insufficient to just
|
||||
approximate the same number of components, for they might be
|
||||
completely different (true components being merged, predicted
|
||||
components splitting others). We really want to capture the
|
||||
correspondence between those labels, which is localised.
|
||||
|
||||
For that we now calculate the label pairs and their counts.
|
||||
Looking at the M,N incidence matrix, we want those counts
|
||||
to be distributed orthogonally (ideally). So we compute a
|
||||
singular value decomposition and compare the sum total of
|
||||
singular values to the sum total of all label counts. The
|
||||
rate of the two determines a measure of congruence.
|
||||
|
||||
Moreover, for the case of artificial boundary segments around
|
||||
regions, optionally introduced by the training extractor to
|
||||
represent segment identity in the loss (and removed at runtime):
|
||||
Reduce this class to background as well.
|
||||
"""
|
||||
def metric(y_true, y_pred):
|
||||
if artificial:
|
||||
# convert artificial border class to background
|
||||
y_true = y_true[:, :, :, :artificial]
|
||||
y_pred = y_pred[:, :, :, :artificial]
|
||||
# [B, H, W, C]
|
||||
l_true = tf.math.argmax(y_true, axis=-1)
|
||||
l_pred = tf.math.argmax(y_pred, axis=-1)
|
||||
# [B, H, W]
|
||||
c_true = tf.cast(connected_components(l_true), tf.int64)
|
||||
c_pred = tf.cast(connected_components(l_pred), tf.int64)
|
||||
# [B, H, W]
|
||||
n_batch = y_true.shape[0]
|
||||
C_true = tf.math.reduce_max(c_true, (1, 2)) + 1
|
||||
C_pred = tf.math.reduce_max(c_pred, (1, 2)) + 1
|
||||
MODULUS = tf.constant(2**22, tf.int64)
|
||||
tf.debugging.assert_less(C_true, MODULUS,
|
||||
message="cannot compare segments: too many connected components in GT")
|
||||
tf.debugging.assert_less(C_pred, MODULUS,
|
||||
message="cannot compare segments: too many connected components in prediction")
|
||||
c_comb = MODULUS * c_pred + c_true
|
||||
tf.debugging.assert_greater_equal(c_comb, tf.constant(0, tf.int64),
|
||||
message="overflow pairing components")
|
||||
# [B, H, W]
|
||||
# tf.unique does not support batch dim, so...
|
||||
results = []
|
||||
for c_comb, C_true, C_pred in zip(
|
||||
tf.unstack(c_comb, num=n_batch),
|
||||
tf.unstack(C_true, num=n_batch),
|
||||
tf.unstack(C_pred, num=n_batch),
|
||||
):
|
||||
prod, _, count = tf.unique_with_counts(tf.reshape(c_comb, (-1,)))
|
||||
# [L]
|
||||
#corr = tf.zeros([C_pred, C_true], tf.int32)
|
||||
#corr[prod // 2**24, prod % 2**24] = count
|
||||
corr = tf.scatter_nd(tf.stack([prod // MODULUS, prod % MODULUS], axis=1),
|
||||
count, (C_pred, C_true))
|
||||
corr = tf.cast(corr, tf.float32)
|
||||
# [Cpred, Ctrue]
|
||||
sgv = tf.linalg.svd(corr, compute_uv=False)
|
||||
results.append(tf.reduce_sum(sgv) / tf.reduce_sum(corr))
|
||||
return 1.0 - tf.reduce_mean(tf.stack(results), 0)
|
||||
# c_true = tf.reshape(c_true, (n_batch, -1))
|
||||
# c_pred = tf.reshape(c_pred, (n_batch, -1))
|
||||
# # [B, H*W]
|
||||
# n_true = tf.math.reduce_max(c_true, axis=1)
|
||||
# n_pred = tf.math.reduce_max(c_pred, axis=1)
|
||||
# # [B]
|
||||
# diff = tf.cast(n_true - n_pred, tf.float32)
|
||||
# return tf.reduce_mean(tf.math.abs(diff) + alpha * diff, axis=-1)
|
||||
|
||||
metric.__name__ = 'nCC'
|
||||
metric._direction = 'down'
|
||||
return metric
|
||||
|
||||
|
|
@ -1,502 +0,0 @@
|
|||
import os
|
||||
|
||||
os.environ['TF_USE_LEGACY_KERAS'] = '1' # avoid Keras 3 after TF 2.15
|
||||
import tensorflow as tf
|
||||
from tensorflow.keras.layers import (
|
||||
Activation,
|
||||
Add,
|
||||
AveragePooling2D,
|
||||
BatchNormalization,
|
||||
Bidirectional,
|
||||
Conv1D,
|
||||
Conv2D,
|
||||
Dense,
|
||||
Dropout,
|
||||
Embedding,
|
||||
Flatten,
|
||||
Input,
|
||||
Layer,
|
||||
LayerNormalization,
|
||||
LSTM,
|
||||
MaxPooling2D,
|
||||
MultiHeadAttention,
|
||||
Reshape,
|
||||
UpSampling2D,
|
||||
ZeroPadding2D,
|
||||
add,
|
||||
concatenate
|
||||
)
|
||||
from tensorflow.keras.models import Model
|
||||
from tensorflow.keras.regularizers import l2
|
||||
from tensorflow.keras.backend import ctc_batch_cost
|
||||
|
||||
from ..patch_encoder import Patches, PatchEncoder
|
||||
|
||||
##mlp_head_units = [512, 256]#[2048, 1024]
|
||||
###projection_dim = 64
|
||||
##transformer_layers = 2#8
|
||||
##num_heads = 1#4
|
||||
RESNET50_WEIGHTS_PATH = './pretrained_model/resnet50_weights_tf_dim_ordering_tf_kernels_notop.h5'
|
||||
RESNET50_WEIGHTS_URL = ('https://github.com/fchollet/deep-learning-models/releases/download/v0.2/'
|
||||
'resnet50_weights_tf_dim_ordering_tf_kernels_notop.h5')
|
||||
|
||||
IMAGE_ORDERING = 'channels_last'
|
||||
MERGE_AXIS = -1
|
||||
|
||||
|
||||
class CTCLayer(Layer):
|
||||
def call(self, y_true, y_pred):
|
||||
batch_len = tf.cast(tf.shape(y_true)[0], dtype="int64")
|
||||
input_length = tf.cast(tf.shape(y_pred)[1], dtype="int64")
|
||||
label_length = tf.cast(tf.shape(y_true)[1], dtype="int64")
|
||||
|
||||
input_length = input_length * tf.ones(shape=(batch_len, 1), dtype="int64")
|
||||
label_length = label_length * tf.ones(shape=(batch_len, 1), dtype="int64")
|
||||
loss = ctc_batch_cost(y_true, y_pred, input_length, label_length)
|
||||
self.add_loss(loss)
|
||||
|
||||
# At test time, just return the computed predictions.
|
||||
return y_pred
|
||||
|
||||
def mlp(x, hidden_units, dropout_rate):
|
||||
for units in hidden_units:
|
||||
x = Dense(units, activation=tf.nn.gelu)(x)
|
||||
x = Dropout(dropout_rate)(x)
|
||||
return x
|
||||
|
||||
def one_side_pad(x):
|
||||
x = ZeroPadding2D(((1, 0), (1, 0)), data_format=IMAGE_ORDERING)(x)
|
||||
return x
|
||||
|
||||
def identity_block(input_tensor, kernel_size, filters, stage, block):
|
||||
"""The identity block is the block that has no conv layer at shortcut.
|
||||
# Arguments
|
||||
input_tensor: input tensor
|
||||
kernel_size: defualt 3, the kernel size of middle conv layer at main path
|
||||
filters: list of integers, the filterss of 3 conv layer at main path
|
||||
stage: integer, current stage label, used for generating layer names
|
||||
block: 'a','b'..., current block label, used for generating layer names
|
||||
# Returns
|
||||
Output tensor for the block.
|
||||
"""
|
||||
filters1, filters2, filters3 = filters
|
||||
|
||||
if IMAGE_ORDERING == 'channels_last':
|
||||
bn_axis = 3
|
||||
else:
|
||||
bn_axis = 1
|
||||
|
||||
conv_name_base = 'res' + str(stage) + block + '_branch'
|
||||
bn_name_base = 'bn' + str(stage) + block + '_branch'
|
||||
|
||||
x = Conv2D(filters1, (1, 1), data_format=IMAGE_ORDERING, name=conv_name_base + '2a')(input_tensor)
|
||||
x = BatchNormalization(axis=bn_axis, name=bn_name_base + '2a')(x)
|
||||
x = Activation('relu')(x)
|
||||
|
||||
x = Conv2D(filters2, kernel_size, data_format=IMAGE_ORDERING,
|
||||
padding='same', name=conv_name_base + '2b')(x)
|
||||
x = BatchNormalization(axis=bn_axis, name=bn_name_base + '2b')(x)
|
||||
x = Activation('relu')(x)
|
||||
|
||||
x = Conv2D(filters3, (1, 1), data_format=IMAGE_ORDERING, name=conv_name_base + '2c')(x)
|
||||
x = BatchNormalization(axis=bn_axis, name=bn_name_base + '2c')(x)
|
||||
|
||||
x = add([x, input_tensor])
|
||||
x = Activation('relu')(x)
|
||||
return x
|
||||
|
||||
|
||||
def conv_block(input_tensor, kernel_size, filters, stage, block, strides=(2, 2)):
|
||||
"""conv_block is the block that has a conv layer at shortcut
|
||||
# Arguments
|
||||
input_tensor: input tensor
|
||||
kernel_size: defualt 3, the kernel size of middle conv layer at main path
|
||||
filters: list of integers, the filterss of 3 conv layer at main path
|
||||
stage: integer, current stage label, used for generating layer names
|
||||
block: 'a','b'..., current block label, used for generating layer names
|
||||
# Returns
|
||||
Output tensor for the block.
|
||||
Note that from stage 3, the first conv layer at main path is with strides=(2,2)
|
||||
And the shortcut should have strides=(2,2) as well
|
||||
"""
|
||||
filters1, filters2, filters3 = filters
|
||||
|
||||
if IMAGE_ORDERING == 'channels_last':
|
||||
bn_axis = 3
|
||||
else:
|
||||
bn_axis = 1
|
||||
|
||||
conv_name_base = 'res' + str(stage) + block + '_branch'
|
||||
bn_name_base = 'bn' + str(stage) + block + '_branch'
|
||||
|
||||
x = Conv2D(filters1, (1, 1), data_format=IMAGE_ORDERING, strides=strides,
|
||||
name=conv_name_base + '2a')(input_tensor)
|
||||
x = BatchNormalization(axis=bn_axis, name=bn_name_base + '2a')(x)
|
||||
x = Activation('relu')(x)
|
||||
|
||||
x = Conv2D(filters2, kernel_size, data_format=IMAGE_ORDERING, padding='same',
|
||||
name=conv_name_base + '2b')(x)
|
||||
x = BatchNormalization(axis=bn_axis, name=bn_name_base + '2b')(x)
|
||||
x = Activation('relu')(x)
|
||||
|
||||
x = Conv2D(filters3, (1, 1), data_format=IMAGE_ORDERING, name=conv_name_base + '2c')(x)
|
||||
x = BatchNormalization(axis=bn_axis, name=bn_name_base + '2c')(x)
|
||||
|
||||
shortcut = Conv2D(filters3, (1, 1), data_format=IMAGE_ORDERING, strides=strides,
|
||||
name=conv_name_base + '1')(input_tensor)
|
||||
shortcut = BatchNormalization(axis=bn_axis, name=bn_name_base + '1')(shortcut)
|
||||
|
||||
x = add([x, shortcut])
|
||||
x = Activation('relu')(x)
|
||||
return x
|
||||
|
||||
def resnet50(inputs, weight_decay=1e-6, pretraining=False):
|
||||
if IMAGE_ORDERING == 'channels_last':
|
||||
bn_axis = 3
|
||||
else:
|
||||
bn_axis = 1
|
||||
|
||||
x = ZeroPadding2D((3, 3), data_format=IMAGE_ORDERING)(inputs)
|
||||
x = Conv2D(64, (7, 7), data_format=IMAGE_ORDERING, strides=(2, 2), kernel_regularizer=l2(weight_decay),
|
||||
name='conv1')(x)
|
||||
f1 = x
|
||||
|
||||
x = BatchNormalization(axis=bn_axis, name='bn_conv1')(x)
|
||||
x = Activation('relu')(x)
|
||||
x = MaxPooling2D((3, 3), data_format=IMAGE_ORDERING, strides=(2, 2))(x)
|
||||
|
||||
x = conv_block(x, 3, [64, 64, 256], stage=2, block='a', strides=(1, 1))
|
||||
x = identity_block(x, 3, [64, 64, 256], stage=2, block='b')
|
||||
x = identity_block(x, 3, [64, 64, 256], stage=2, block='c')
|
||||
f2 = one_side_pad(x)
|
||||
|
||||
x = conv_block(x, 3, [128, 128, 512], stage=3, block='a')
|
||||
x = identity_block(x, 3, [128, 128, 512], stage=3, block='b')
|
||||
x = identity_block(x, 3, [128, 128, 512], stage=3, block='c')
|
||||
x = identity_block(x, 3, [128, 128, 512], stage=3, block='d')
|
||||
f3 = x
|
||||
|
||||
x = conv_block(x, 3, [256, 256, 1024], stage=4, block='a')
|
||||
x = identity_block(x, 3, [256, 256, 1024], stage=4, block='b')
|
||||
x = identity_block(x, 3, [256, 256, 1024], stage=4, block='c')
|
||||
x = identity_block(x, 3, [256, 256, 1024], stage=4, block='d')
|
||||
x = identity_block(x, 3, [256, 256, 1024], stage=4, block='e')
|
||||
x = identity_block(x, 3, [256, 256, 1024], stage=4, block='f')
|
||||
f4 = x
|
||||
|
||||
x = conv_block(x, 3, [512, 512, 2048], stage=5, block='a')
|
||||
x = identity_block(x, 3, [512, 512, 2048], stage=5, block='b')
|
||||
x = identity_block(x, 3, [512, 512, 2048], stage=5, block='c')
|
||||
f5 = x
|
||||
|
||||
if pretraining:
|
||||
model = Model(inputs, x).load_weights(RESNET50_WEIGHTS_PATH)
|
||||
|
||||
return f1, f2, f3, f4, f5
|
||||
|
||||
def unet_decoder(img, f1, f2, f3, f4, f5, n_classes, light=False, task="segmentation", weight_decay=1e-6):
|
||||
if IMAGE_ORDERING == 'channels_last':
|
||||
bn_axis = 3
|
||||
else:
|
||||
bn_axis = 1
|
||||
|
||||
o = Conv2D(512 if light else 1024, (1, 1), padding='same',
|
||||
data_format=IMAGE_ORDERING, kernel_regularizer=l2(weight_decay))(f5)
|
||||
o = BatchNormalization(axis=bn_axis)(o)
|
||||
o = Activation('relu')(o)
|
||||
|
||||
if light:
|
||||
f4 = Conv2D(512, (1, 1), padding='same',
|
||||
data_format=IMAGE_ORDERING, kernel_regularizer=l2(weight_decay))(f4)
|
||||
f4 = BatchNormalization(axis=bn_axis)(f4)
|
||||
f4 = Activation('relu')(f4)
|
||||
|
||||
o = UpSampling2D((2, 2), data_format=IMAGE_ORDERING, interpolation="bilinear")(o)
|
||||
o = concatenate([o, f4], axis=MERGE_AXIS)
|
||||
o = ZeroPadding2D((1, 1), data_format=IMAGE_ORDERING)(o)
|
||||
o = Conv2D(512, (3, 3), padding='valid',
|
||||
data_format=IMAGE_ORDERING, kernel_regularizer=l2(weight_decay))(o)
|
||||
o = BatchNormalization(axis=bn_axis)(o)
|
||||
o = Activation('relu')(o)
|
||||
|
||||
o = UpSampling2D((2, 2), data_format=IMAGE_ORDERING, interpolation="bilinear")(o)
|
||||
o = concatenate([o, f3], axis=MERGE_AXIS)
|
||||
o = ZeroPadding2D((1, 1), data_format=IMAGE_ORDERING)(o)
|
||||
o = Conv2D(256, (3, 3), padding='valid',
|
||||
data_format=IMAGE_ORDERING, kernel_regularizer=l2(weight_decay))(o)
|
||||
o = BatchNormalization(axis=bn_axis)(o)
|
||||
o = Activation('relu')(o)
|
||||
|
||||
o = UpSampling2D((2, 2), data_format=IMAGE_ORDERING, interpolation="bilinear")(o)
|
||||
o = concatenate([o, f2], axis=MERGE_AXIS)
|
||||
o = ZeroPadding2D((1, 1), data_format=IMAGE_ORDERING)(o)
|
||||
o = Conv2D(128, (3, 3), padding='valid',
|
||||
data_format=IMAGE_ORDERING, kernel_regularizer=l2(weight_decay))(o)
|
||||
o = BatchNormalization(axis=bn_axis)(o)
|
||||
o = Activation('relu')(o)
|
||||
|
||||
o = UpSampling2D((2, 2), data_format=IMAGE_ORDERING, interpolation="bilinear")(o)
|
||||
o = concatenate([o, f1], axis=MERGE_AXIS)
|
||||
o = ZeroPadding2D((1, 1), data_format=IMAGE_ORDERING)(o)
|
||||
o = Conv2D(64, (3, 3), padding='valid',
|
||||
data_format=IMAGE_ORDERING, kernel_regularizer=l2(weight_decay))(o)
|
||||
o = BatchNormalization(axis=bn_axis)(o)
|
||||
o = Activation('relu')(o)
|
||||
|
||||
o = UpSampling2D((2, 2), data_format=IMAGE_ORDERING, interpolation="bilinear")(o)
|
||||
o = concatenate([o, img], axis=MERGE_AXIS)
|
||||
o = ZeroPadding2D((1, 1), data_format=IMAGE_ORDERING)(o)
|
||||
o = Conv2D(32, (3, 3), padding='valid',
|
||||
data_format=IMAGE_ORDERING, kernel_regularizer=l2(weight_decay))(o)
|
||||
o = BatchNormalization(axis=bn_axis)(o)
|
||||
o = Activation('relu')(o)
|
||||
|
||||
o = Conv2D(n_classes, (1, 1), padding='same',
|
||||
data_format=IMAGE_ORDERING, kernel_regularizer=l2(weight_decay))(o)
|
||||
if task == "segmentation":
|
||||
o = BatchNormalization(axis=bn_axis)(o)
|
||||
o = Activation('softmax')(o)
|
||||
else:
|
||||
o = Activation('sigmoid')(o)
|
||||
|
||||
return Model(img, o)
|
||||
|
||||
def resnet50_unet_light(n_classes, input_height=224, input_width=224, task="segmentation", weight_decay=1e-6, pretraining=False):
|
||||
assert input_height % 32 == 0
|
||||
assert input_width % 32 == 0
|
||||
|
||||
img_input = Input(shape=(input_height, input_width, 3))
|
||||
|
||||
features = resnet50(img_input, weight_decay=weight_decay, pretraining=pretraining)
|
||||
|
||||
return unet_decoder(img_input, *features, n_classes, light=True, task=task, weight_decay=weight_decay)
|
||||
|
||||
def resnet50_unet(n_classes, input_height=224, input_width=224, task="segmentation", weight_decay=1e-6, pretraining=False):
|
||||
assert input_height % 32 == 0
|
||||
assert input_width % 32 == 0
|
||||
|
||||
img_input = Input(shape=(input_height, input_width, 3))
|
||||
|
||||
features = resnet50(img_input, weight_decay=weight_decay, pretraining=pretraining)
|
||||
|
||||
return unet_decoder(img_input, *features, n_classes, light=False, task=task, weight_decay=weight_decay)
|
||||
|
||||
def transformer_block(img,
|
||||
num_patches,
|
||||
patchsize_x,
|
||||
patchsize_y,
|
||||
mlp_head_units,
|
||||
n_layers,
|
||||
num_heads,
|
||||
projection_dim):
|
||||
patches = Patches(patchsize_x, patchsize_y)(img)
|
||||
# Encode patches.
|
||||
encoded_patches = PatchEncoder(num_patches, projection_dim)(patches)
|
||||
|
||||
for _ in range(n_layers):
|
||||
# Layer normalization 1.
|
||||
x1 = LayerNormalization(epsilon=1e-6)(encoded_patches)
|
||||
# Create a multi-head attention layer.
|
||||
attention_output = MultiHeadAttention(num_heads=num_heads,
|
||||
key_dim=projection_dim,
|
||||
dropout=0.1)(x1, x1)
|
||||
# Skip connection 1.
|
||||
x2 = Add()([attention_output, encoded_patches])
|
||||
# Layer normalization 2.
|
||||
x3 = LayerNormalization(epsilon=1e-6)(x2)
|
||||
# MLP.
|
||||
x3 = mlp(x3, hidden_units=mlp_head_units, dropout_rate=0.1)
|
||||
# Skip connection 2.
|
||||
encoded_patches = Add()([x3, x2])
|
||||
|
||||
encoded_patches = tf.reshape(encoded_patches,
|
||||
[-1,
|
||||
img.shape[1],
|
||||
img.shape[2],
|
||||
projection_dim // (patchsize_x * patchsize_y)])
|
||||
return encoded_patches
|
||||
|
||||
def vit_resnet50_unet(num_patches,
|
||||
n_classes,
|
||||
transformer_patchsize_x,
|
||||
transformer_patchsize_y,
|
||||
transformer_mlp_head_units=None,
|
||||
transformer_layers=8,
|
||||
transformer_num_heads=4,
|
||||
transformer_projection_dim=64,
|
||||
input_height=224,
|
||||
input_width=224,
|
||||
task="segmentation",
|
||||
weight_decay=1e-6,
|
||||
pretraining=False):
|
||||
if transformer_mlp_head_units is None:
|
||||
transformer_mlp_head_units = [128, 64]
|
||||
inputs = Input(shape=(input_height, input_width, 3))
|
||||
|
||||
features = list(resnet50(inputs, weight_decay=weight_decay, pretraining=pretraining))
|
||||
|
||||
features[-1] = transformer_block(features[-1],
|
||||
num_patches,
|
||||
transformer_patchsize_x,
|
||||
transformer_patchsize_y,
|
||||
transformer_mlp_head_units,
|
||||
transformer_layers,
|
||||
transformer_num_heads,
|
||||
transformer_projection_dim)
|
||||
|
||||
return unet_decoder(inputs, *features, n_classes, task=task, weight_decay=weight_decay)
|
||||
|
||||
def vit_resnet50_unet_transformer_before_cnn(num_patches,
|
||||
n_classes,
|
||||
transformer_patchsize_x,
|
||||
transformer_patchsize_y,
|
||||
transformer_mlp_head_units=None,
|
||||
transformer_layers=8,
|
||||
transformer_num_heads=4,
|
||||
transformer_projection_dim=64,
|
||||
input_height=224,
|
||||
input_width=224,
|
||||
task="segmentation",
|
||||
weight_decay=1e-6,
|
||||
pretraining=False):
|
||||
if transformer_mlp_head_units is None:
|
||||
transformer_mlp_head_units = [128, 64]
|
||||
inputs = Input(shape=(input_height, input_width, 3))
|
||||
|
||||
encoded_patches = transformer_block(inputs,
|
||||
num_patches,
|
||||
transformer_patchsize_x,
|
||||
transformer_patchsize_y,
|
||||
transformer_mlp_head_units,
|
||||
transformer_layers,
|
||||
transformer_num_heads,
|
||||
transformer_projection_dim)
|
||||
encoded_patches = Conv2D(3, (1, 1), padding='same',
|
||||
data_format=IMAGE_ORDERING, kernel_regularizer=l2(weight_decay),
|
||||
name='convinput')(encoded_patches)
|
||||
|
||||
features = resnet50(encoded_patches, weight_decay=weight_decay, pretraining=pretraining)
|
||||
|
||||
return unet_decoder(inputs, *features, n_classes, task=task, weight_decay=weight_decay)
|
||||
|
||||
def resnet50_classifier(n_classes,input_height=224,input_width=224,weight_decay=1e-6,pretraining=False):
|
||||
include_top=True
|
||||
assert input_height%32 == 0
|
||||
assert input_width%32 == 0
|
||||
|
||||
|
||||
img_input = Input(shape=(input_height,input_width , 3 ))
|
||||
|
||||
_, _, _, _, x = resnet50(img_input, weight_decay, pretraining)
|
||||
|
||||
x = AveragePooling2D((7, 7), name='avg_pool')(x)
|
||||
x = Flatten()(x)
|
||||
|
||||
##
|
||||
x = Dense(256, activation='relu', name='fc512')(x)
|
||||
x=Dropout(0.2)(x)
|
||||
##
|
||||
x = Dense(n_classes, activation='softmax', name='fc1000')(x)
|
||||
model = Model(img_input, x)
|
||||
|
||||
return model
|
||||
|
||||
def machine_based_reading_order_model(n_classes,input_height=224,input_width=224,weight_decay=1e-6,pretraining=False):
|
||||
assert input_height%32 == 0
|
||||
assert input_width%32 == 0
|
||||
|
||||
img_input = Input(shape=(input_height,input_width , 3 ))
|
||||
|
||||
_, _, _, _, x = resnet50(img_input, weight_decay, pretraining)
|
||||
|
||||
x = AveragePooling2D((7, 7), name='avg_pool1')(x)
|
||||
flattened = Flatten()(x)
|
||||
|
||||
o = Dense(256, activation='relu', name='fc512')(flattened)
|
||||
o=Dropout(0.2)(o)
|
||||
|
||||
o = Dense(256, activation='relu', name='fc512a')(o)
|
||||
o=Dropout(0.2)(o)
|
||||
|
||||
o = Dense(n_classes, activation='sigmoid', name='fc1000')(o)
|
||||
model = Model(img_input , o)
|
||||
|
||||
return model
|
||||
|
||||
def cnn_rnn_ocr_model(image_height=None, image_width=None, n_classes=None, max_seq=None):
|
||||
input_img = Input(shape=(image_height, image_width, 3), name="image")
|
||||
labels = Input(name="label", shape=(None,))
|
||||
|
||||
x = Conv2D(64,kernel_size=(3,3),padding="same")(input_img)
|
||||
x = BatchNormalization(name="bn1")(x)
|
||||
x = Activation("relu", name="relu1")(x)
|
||||
x = Conv2D(64,kernel_size=(3,3),padding="same")(x)
|
||||
x = BatchNormalization(name="bn2")(x)
|
||||
x = Activation("relu", name="relu2")(x)
|
||||
x = MaxPooling2D(pool_size=(1,2),strides=(1,2))(x)
|
||||
|
||||
x = Conv2D(128,kernel_size=(3,3),padding="same")(x)
|
||||
x = BatchNormalization(name="bn3")(x)
|
||||
x = Activation("relu", name="relu3")(x)
|
||||
x = Conv2D(128,kernel_size=(3,3),padding="same")(x)
|
||||
x = BatchNormalization(name="bn4")(x)
|
||||
x = Activation("relu", name="relu4")(x)
|
||||
x = MaxPooling2D(pool_size=(1,2),strides=(1,2))(x)
|
||||
|
||||
x = Conv2D(256,kernel_size=(3,3),padding="same")(x)
|
||||
x = BatchNormalization(name="bn5")(x)
|
||||
x = Activation("relu", name="relu5")(x)
|
||||
x = Conv2D(256,kernel_size=(3,3),padding="same")(x)
|
||||
x = BatchNormalization(name="bn6")(x)
|
||||
x = Activation("relu", name="relu6")(x)
|
||||
x = MaxPooling2D(pool_size=(2,2),strides=(2,2))(x)
|
||||
|
||||
x = Conv2D(image_width,kernel_size=(3,3),padding="same")(x)
|
||||
x = BatchNormalization(name="bn7")(x)
|
||||
x = Activation("relu", name="relu7")(x)
|
||||
x = Conv2D(image_width,kernel_size=(16,1))(x)
|
||||
x = BatchNormalization(name="bn8")(x)
|
||||
x = Activation("relu", name="relu8")(x)
|
||||
x2d = MaxPooling2D(pool_size=(1,2),strides=(1,2))(x)
|
||||
x4d = MaxPooling2D(pool_size=(1,2),strides=(1,2))(x2d)
|
||||
|
||||
|
||||
new_shape = (x.shape[1]*x.shape[2], x.shape[3])
|
||||
new_shape2 = (x2d.shape[1]*x2d.shape[2], x2d.shape[3])
|
||||
new_shape4 = (x4d.shape[1]*x4d.shape[2], x4d.shape[3])
|
||||
|
||||
x = Reshape(target_shape=new_shape, name="reshape")(x)
|
||||
x2d = Reshape(target_shape=new_shape2, name="reshape2")(x2d)
|
||||
x4d = Reshape(target_shape=new_shape4, name="reshape4")(x4d)
|
||||
|
||||
xrnnorg = Bidirectional(LSTM(image_width, return_sequences=True, dropout=0.25))(x)
|
||||
xrnn2d = Bidirectional(LSTM(image_width, return_sequences=True, dropout=0.25))(x2d)
|
||||
xrnn4d = Bidirectional(LSTM(image_width, return_sequences=True, dropout=0.25))(x4d)
|
||||
|
||||
xrnn2d = Reshape(target_shape=(1, xrnn2d.shape[1], xrnn2d.shape[2]), name="reshape6")(xrnn2d)
|
||||
xrnn4d = Reshape(target_shape=(1, xrnn4d.shape[1], xrnn4d.shape[2]), name="reshape8")(xrnn4d)
|
||||
|
||||
|
||||
xrnn2dup = UpSampling2D(size=(1, 2), interpolation="nearest")(xrnn2d)
|
||||
xrnn4dup = UpSampling2D(size=(1, 4), interpolation="nearest")(xrnn4d)
|
||||
|
||||
xrnn2dup = Reshape(target_shape=(xrnn2dup.shape[2], xrnn2dup.shape[3]), name="reshape10")(xrnn2dup)
|
||||
xrnn4dup = Reshape(target_shape=(xrnn4dup.shape[2], xrnn4dup.shape[3]), name="reshape12")(xrnn4dup)
|
||||
|
||||
addition = Add()([xrnnorg, xrnn2dup, xrnn4dup])
|
||||
|
||||
addition_rnn = Bidirectional(LSTM(image_width, return_sequences=True, dropout=0.25))(addition)
|
||||
|
||||
out = Conv1D(max_seq, 1, data_format="channels_first")(addition_rnn)
|
||||
out = BatchNormalization(name="bn9")(out)
|
||||
out = Activation("relu", name="relu9")(out)
|
||||
#out = Conv1D(n_classes, 1, activation='relu', data_format="channels_last")(out)
|
||||
|
||||
out = Dense(n_classes, activation="softmax", name="dense2")(out)
|
||||
|
||||
# Add CTC layer for calculating CTC loss at each step.
|
||||
output = CTCLayer(name="ctc_loss")(labels, out)
|
||||
|
||||
model = Model(inputs=(input_img, labels), outputs=output, name="handwriting_recognizer")
|
||||
|
||||
return model
|
||||
|
|
@ -1,48 +0,0 @@
|
|||
SHELL = bash -e
|
||||
|
||||
MODELS_SRC = models_eynollah
|
||||
MODELS_DST = reloaded/models_eynollah
|
||||
|
||||
|
||||
# $(MODELS_DST)/eynollah-binarization_20210425 \
|
||||
# $(MODELS_DST)/eynollah-column-classifier_20210425 \
|
||||
# $(MODELS_DST)/eynollah-enhancement_20210425 \
|
||||
# $(MODELS_DST)/eynollah-main-regions-aug-rotation_20210425 \
|
||||
# $(MODELS_DST)/eynollah-main-regions-aug-scaling_20210425 \
|
||||
# $(MODELS_DST)/eynollah-main-regions-ensembled_20210425 \
|
||||
# $(MODELS_DST)/eynollah-main-regions_20220314 \
|
||||
# $(MODELS_DST)/eynollah-main-regions_20231127_672_org_ens_11_13_16_17_18 \
|
||||
# $(MODELS_DST)/eynollah-tables_20210319 \
|
||||
# $(MODELS_DST)/model_eynollah_ocr_cnnrnn_20250930 \
|
||||
|
||||
RELOADABLE_MODELS = \
|
||||
$(MODELS_DST)/model_eynollah_page_extraction_20250915 \
|
||||
$(MODELS_DST)/model_eynollah_reading_order_20250824 \
|
||||
$(MODELS_DST)/modelens_e_l_all_sp_0_1_2_3_4_171024 \
|
||||
$(MODELS_DST)/modelens_full_lay_1__4_3_091124 \
|
||||
$(MODELS_DST)/modelens_table_0t4_201124 \
|
||||
$(MODELS_DST)/modelens_textline_0_1__2_4_16092024
|
||||
|
||||
all: $(RELOADABLE_MODELS)
|
||||
|
||||
$(MODELS_DST)/%: $(MODELS_SRC)/%
|
||||
mkdir -p $@
|
||||
test -e $</config.json || exit 1
|
||||
eynollah-training train --force \
|
||||
with $</config.json \
|
||||
reload_weights=True \
|
||||
continue_training=False \
|
||||
dir_output=$(dir $@) \
|
||||
dir_of_start_model=$< \
|
||||
2>&1 | tee $(notdir $<).log
|
||||
cp $</config.json $@/config.json
|
||||
|
||||
compare:
|
||||
for i in `find $(MODELS_DST) -mindepth 2`;do \
|
||||
n=$(MODELS_SRC)$${i#$(MODELS_DST)}; \
|
||||
du -bs $$n $$i ; \
|
||||
done
|
||||
|
||||
|
||||
clear:
|
||||
rm -rf $(MODELS_DST)
|
||||
|
|
@ -1,883 +0,0 @@
|
|||
import os
|
||||
import sys
|
||||
import io
|
||||
import json
|
||||
|
||||
from tqdm import tqdm
|
||||
import requests
|
||||
|
||||
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
|
||||
os.environ['TF_USE_LEGACY_KERAS'] = '1' # avoid Keras 3 after TF 2.15
|
||||
import tensorflow as tf
|
||||
from tensorflow.keras.optimizers import SGD, Adam
|
||||
from tensorflow.keras.metrics import MeanIoU, F1Score
|
||||
from tensorflow.keras.models import load_model
|
||||
from tensorflow.keras.callbacks import ModelCheckpoint, TensorBoard, EarlyStopping
|
||||
from tensorflow.keras.layers import StringLookup
|
||||
from tensorflow.keras.utils import image_dataset_from_directory
|
||||
from tensorflow.keras.backend import one_hot
|
||||
from sacred import Experiment
|
||||
from sacred.config import create_captured_function
|
||||
|
||||
import numpy as np
|
||||
import cv2
|
||||
from matplotlib import pyplot as plt # for plot_confusion_matrix
|
||||
|
||||
from .metrics import (
|
||||
soft_dice_loss,
|
||||
weighted_categorical_crossentropy,
|
||||
get as get_metric,
|
||||
metrics_superposition,
|
||||
ConfusionMatrix,
|
||||
connected_components_loss,
|
||||
)
|
||||
from .models import (
|
||||
PatchEncoder,
|
||||
Patches,
|
||||
machine_based_reading_order_model,
|
||||
resnet50_classifier,
|
||||
resnet50_unet,
|
||||
vit_resnet50_unet,
|
||||
vit_resnet50_unet_transformer_before_cnn,
|
||||
cnn_rnn_ocr_model,
|
||||
RESNET50_WEIGHTS_PATH,
|
||||
RESNET50_WEIGHTS_URL
|
||||
)
|
||||
from .utils import (
|
||||
generate_arrays_from_folder_reading_order,
|
||||
get_one_hot,
|
||||
preprocess_imgs,
|
||||
)
|
||||
from .weights_ensembling import run_ensembling
|
||||
|
||||
|
||||
class SaveWeightsAfterSteps(ModelCheckpoint):
|
||||
def __init__(self, save_interval, save_path, _config, **kwargs):
|
||||
if save_interval:
|
||||
# batches
|
||||
super().__init__(
|
||||
os.path.join(save_path, "model_step_{batch:04d}"),
|
||||
save_freq=save_interval,
|
||||
verbose=1,
|
||||
**kwargs)
|
||||
else:
|
||||
super().__init__(
|
||||
os.path.join(save_path, "model_{epoch:02d}"),
|
||||
save_freq="epoch",
|
||||
verbose=1,
|
||||
**kwargs)
|
||||
self._config = _config
|
||||
|
||||
# overwrite tf-keras (Keras 2) implementation to get our _config JSON in
|
||||
def _save_handler(self, filepath):
|
||||
super()._save_handler(filepath)
|
||||
with open(os.path.join(filepath, "config.json"), "w") as fp:
|
||||
json.dump(self._config, fp) # encode dict into JSON
|
||||
|
||||
def configuration():
|
||||
try:
|
||||
for device in tf.config.list_physical_devices('GPU'):
|
||||
tf.config.experimental.set_memory_growth(device, True)
|
||||
#tf.keras.mixed_precision.set_global_policy('mixed_float16')
|
||||
#tf.keras.backend.set_epsilon(1e-4) # avoid NaN from smaller defaults
|
||||
except:
|
||||
print("no GPU device available", file=sys.stderr)
|
||||
|
||||
def plot_layout_tf(in_: tf.Tensor, out:tf.Tensor) -> tf.Tensor:
|
||||
"""
|
||||
Implements training.inference.SBBPredict.visualize_model_output for TF
|
||||
(effectively plotting the layout segmentation map on the input image).
|
||||
|
||||
In doing so, also converts:
|
||||
- from Eynollah's BGR/float on the input side
|
||||
- to std RGB/int format on the output side
|
||||
"""
|
||||
# in_: [B, H, W, 3] (BGR float)
|
||||
image = in_[..., ::-1] * 255
|
||||
# out: [B, H, W, C]
|
||||
lab = tf.math.argmax(out, axis=-1)
|
||||
# lab: [B, H, W]
|
||||
colors = tf.constant([[255, 255, 255],
|
||||
[255, 0, 0],
|
||||
[255, 125, 0],
|
||||
[255, 0, 125],
|
||||
[125, 125, 125],
|
||||
[125, 125, 0],
|
||||
[0, 125, 255],
|
||||
[0, 125, 0],
|
||||
[125, 125, 125],
|
||||
[0, 125, 255],
|
||||
[125, 0, 125],
|
||||
[0, 255, 0],
|
||||
[0, 0, 255],
|
||||
[0, 255, 255],
|
||||
[255, 125, 125],
|
||||
[255, 0, 255]])
|
||||
layout = tf.gather(colors, lab)
|
||||
# layout: [B, H, W, 3]
|
||||
image = tf.cast(image, tf.float32)
|
||||
layout = tf.cast(layout, tf.float32)
|
||||
#weighted = image * 0.5 + layout * 0.1 (too dark)
|
||||
weighted = image * 0.9 + layout * 0.1
|
||||
return tf.cast(weighted, tf.uint8)
|
||||
|
||||
def plot_confusion_matrix(cm, name="Confusion Matrix"):
|
||||
"""
|
||||
Plot the confusion matrix with matplotlib and tensorflow
|
||||
"""
|
||||
fig, ax = plt.subplots(figsize=(10, 8), dpi=300)
|
||||
im = ax.imshow(cm, vmin=0.0, vmax=1.0, interpolation='nearest', cmap=plt.cm.Blues)
|
||||
ax.figure.colorbar(im, ax=ax)
|
||||
ax.set(xticks=np.arange(cm.shape[1]),
|
||||
yticks=np.arange(cm.shape[0]),
|
||||
xlim=[-0.5, cm.shape[1] - 0.5],
|
||||
ylim=[-0.5, cm.shape[0] - 0.5],
|
||||
#xticklabels=labels,
|
||||
#yticklabels=labels,
|
||||
title=name,
|
||||
ylabel='True class',
|
||||
xlabel='Predicted class')
|
||||
# Loop over data dimensions and create text annotations.
|
||||
thresh = cm.max() / 2.
|
||||
for i in range(cm.shape[0]):
|
||||
for j in range(cm.shape[1]):
|
||||
ax.text(j, i, format(cm[i, j], ".2f"),
|
||||
ha="center", va="center",
|
||||
color="white" if cm[i, j] > thresh else "black")
|
||||
fig.tight_layout()
|
||||
# convert to PNG
|
||||
buf = io.BytesIO()
|
||||
fig.savefig(buf, format='png')
|
||||
plt.close(fig)
|
||||
buf.seek(0)
|
||||
# Convert PNG buffer to TF image
|
||||
image = tf.image.decode_png(buf.getvalue(), channels=4)
|
||||
# Add the batch dimension
|
||||
image = tf.expand_dims(image, 0)
|
||||
return image
|
||||
|
||||
# plot predictions on train and test set during every epoch
|
||||
class TensorBoardPlotter(TensorBoard):
|
||||
def __init__(self, plot_freqs, *args, **kwargs):
|
||||
super().__init__(*args, **kwargs)
|
||||
self.model_call = None
|
||||
self.plot_frequency_train, self.plot_frequency_val = plot_freqs
|
||||
def on_epoch_begin(self, epoch, logs=None):
|
||||
super().on_epoch_begin(epoch, logs=logs)
|
||||
# override the model's call(), so we don't have to invest extra cycles
|
||||
# to predict our samples (plotting itself can be neglected)
|
||||
self.model_call = self.model.call
|
||||
def new_call(inputs, **kwargs):
|
||||
outputs = self.model_call(inputs, **kwargs)
|
||||
images = plot_layout_tf(inputs, outputs)
|
||||
self.plot(images, training=kwargs.get('training', None), epoch=epoch)
|
||||
with tf.control_dependencies(None):
|
||||
return outputs
|
||||
self.model.call = new_call
|
||||
# force rebuild of tf.function (so Python binding for epoch gets re-evaluated)
|
||||
self.model.train_function = self.model.make_train_function(True)
|
||||
self.model.test_function = self.model.make_test_function(True)
|
||||
def on_epoch_end(self, epoch, logs=None):
|
||||
# re-instate (so ModelCheckpoint does not see our override call)
|
||||
self.model.call = self.model_call
|
||||
super().on_epoch_end(epoch, logs=logs)
|
||||
def plot(self, images, training=None, epoch=0):
|
||||
if training:
|
||||
writer = self._train_writer
|
||||
freq = self.plot_frequency_train
|
||||
mode, step = "train", self._train_step.value()
|
||||
else:
|
||||
writer = self._val_writer
|
||||
freq = self.plot_frequency_val
|
||||
mode, step = "test", self._val_step.value()
|
||||
# skip most samples, because TF's EncodePNG is so costly,
|
||||
# and now ends up in the middle of our pipeline, thus causing stalls
|
||||
# (cannot use max_outputs, as batch size may be too small)
|
||||
if not tf.cast(step % freq, tf.bool):
|
||||
with writer.as_default():
|
||||
# used to be family kwarg for tf.summary.image name prefix
|
||||
family = "epoch_%03d/" % (1 + epoch)
|
||||
name = family + mode
|
||||
tf.summary.image(name, images, step=step, max_outputs=len(images))
|
||||
def on_train_batch_end(self, batch, logs=None):
|
||||
if logs is not None:
|
||||
logs = dict(logs)
|
||||
# cannot be logged as scalar:
|
||||
logs.pop('confusion_matrix', None)
|
||||
super().on_train_batch_end(batch, logs)
|
||||
def on_test_end(self, logs=None):
|
||||
if logs is not None:
|
||||
logs = dict(logs)
|
||||
# cannot be logged as scalar:
|
||||
logs.pop('confusion_matrix', None)
|
||||
super().on_test_end(logs)
|
||||
def _log_epoch_metrics(self, epoch, logs):
|
||||
if not logs:
|
||||
return
|
||||
logs = dict(logs)
|
||||
# cannot be logged as scalar:
|
||||
train_matrix = logs.pop('confusion_matrix', None)
|
||||
val_matrix = logs.pop('val_confusion_matrix', None)
|
||||
super()._log_epoch_metrics(epoch, logs)
|
||||
# now plot confusion_matrix
|
||||
with tf.summary.record_if(True):
|
||||
if train_matrix is not None:
|
||||
train_image = plot_confusion_matrix(train_matrix)
|
||||
with self._train_writer.as_default():
|
||||
tf.summary.image("confusion_matrix", train_image, step=epoch)
|
||||
if val_matrix is not None:
|
||||
val_image = plot_confusion_matrix(val_matrix)
|
||||
with self._val_writer.as_default():
|
||||
tf.summary.image("confusion_matrix", val_image, step=epoch)
|
||||
|
||||
def get_dirs_or_files(input_data):
|
||||
image_input, labels_input = os.path.join(input_data, 'images/'), os.path.join(input_data, 'labels/')
|
||||
if os.path.isdir(input_data):
|
||||
# Check if training dir exists
|
||||
assert os.path.isdir(image_input), "{} is not a directory".format(image_input)
|
||||
assert os.path.isdir(labels_input), "{} is not a directory".format(labels_input)
|
||||
return image_input, labels_input
|
||||
|
||||
def download_file(url, path):
|
||||
with open(path, 'wb') as f:
|
||||
with requests.get(url, stream=True) as r:
|
||||
r.raise_for_status()
|
||||
for data in r.iter_content(chunk_size=4096):
|
||||
f.write(data)
|
||||
|
||||
ex = Experiment(save_git_info=False)
|
||||
|
||||
|
||||
@ex.config
|
||||
def config_params():
|
||||
task = "segmentation" # This parameter defines task of model which can be segmentation, enhancement or classification.
|
||||
if task in ["segmentation", "binarization", "enhancement"]:
|
||||
backbone_type = "nontransformer" # Type of image feature map network backbone. Either a vision transformer alongside a CNN we call "transformer", or only a CNN which we call "nontransformer"
|
||||
if backbone_type == "transformer":
|
||||
transformer_patchsize_x = None # Patch size of vision transformer patches in x direction.
|
||||
transformer_patchsize_y = None # Patch size of vision transformer patches in y direction.
|
||||
transformer_num_patches_xy = None # Number of patches for vision transformer in x and y direction respectively.
|
||||
transformer_projection_dim = 64 # Transformer projection dimension. Default value is 64.
|
||||
transformer_mlp_head_units = [128, 64] # Transformer Multilayer Perceptron (MLP) head units. Default value is [128, 64]
|
||||
transformer_layers = 8 # transformer layers. Default value is 8.
|
||||
transformer_num_heads = 4 # Transformer number of heads. Default value is 4.
|
||||
transformer_cnn_first = True # We have two types of vision transformers: either the CNN is applied first, followed by the transformer, or reversed.
|
||||
n_classes = None # Number of classes. In the case of binary classification this should be 2.
|
||||
n_epochs = 1 # Number of epochs to train.
|
||||
n_batch = 1 # Number of images per batch at each iteration. (Try as large as fits on VRAM.)
|
||||
if task == 'cnn-rnn-ocr':
|
||||
max_len = None # Maximum sequence length (characters per line) for OCR output.
|
||||
characters_txt_file = None # Path of JSON file defining character set needed of OCR model.
|
||||
input_height = 224 * 1 # Height of model's input in pixels.
|
||||
input_width = 224 * 1 # Width of model's input in pixels.
|
||||
weight_decay = 1e-6 # Weight decay of l2 regularization of model layers.
|
||||
learning_rate = 1e-4 # Set the learning rate.
|
||||
if task in ["segmentation", "binarization"]:
|
||||
is_loss_soft_dice = False # Use soft dice as loss function. When set to true, "weighted_loss" must be false.
|
||||
weighted_loss = False # Use weighted categorical cross entropy as loss fucntion. When set to true, "is_loss_soft_dice" must be false.
|
||||
add_ncc_loss = 0 # Add regression loss for number of connected components. When non-zero, use this as weight for the nCC term.
|
||||
elif task == "classification":
|
||||
f1_threshold_classification = None # This threshold is used to consider models with an evaluation f1 scores bigger than it. The selected model weights undergo a weights ensembling. And avreage ensembled model will be written to output.
|
||||
classification_classes_name = None # Dictionary of classification classes names.
|
||||
patches = False # Divides input image into smaller patches (input size of the model) when set to true. For the model to see the full image, like page extraction, set this to false.
|
||||
augmentation = False # To apply any kind of augmentation, this parameter must be set to true.
|
||||
if augmentation:
|
||||
flip_aug = False # Whether different types of flipping will be applied to the image. Requires "flip_index" setting.
|
||||
blur_aug = False # Whether images will be blurred. Requires "blur_k" setting.
|
||||
if blur_aug:
|
||||
blur_k = None # Method of blurring (gauss, median or blur).
|
||||
padding_white = False # If true, white padding will be applied to the image.
|
||||
if padding_white and task == 'cnn-rnn-ocr':
|
||||
white_padds = None # List of padding sizes.
|
||||
padd_colors = None # List of padding colors, but only "white" or "black" or both.
|
||||
padding_black = False # If true, black padding will be applied to the image.
|
||||
scaling = False # Whether images will be scaled up or down. Requires "scales" setting.
|
||||
scaling_bluring = False # Whether a combination of scaling and blurring will be applied to the image.
|
||||
scaling_binarization = False # Whether a combination of scaling and binarization will be applied to the image.
|
||||
scaling_brightness = False # Whether a combination of scaling and brightening will be applied to the image.
|
||||
scaling_flip = False # Whether a combination of scaling and flipping will be applied to the image.
|
||||
if scaling or scaling_brightness or scaling_bluring or scaling_binarization or scaling_flip:
|
||||
scales = None # Scale patches for augmentation.
|
||||
if flip_aug or scaling_flip:
|
||||
flip_index = None # List of codes (as in cv2.flip) for flip augmentation.
|
||||
shifting = False
|
||||
brightening = False # Whether images will be brightened. Requires "brightness" setting.
|
||||
if brightening:
|
||||
brightness = None # List of intensity factors for brightening.
|
||||
binarization = False # Whether binary images will be used, too. (Will use Otsu thresholding unless supplying precomputed images in "dir_img_bin".)
|
||||
if binarization:
|
||||
dir_img_bin = None # Directory of training dataset subdirectory of binarized images
|
||||
add_red_textlines = False
|
||||
adding_rgb_background = False # Whether texture images will be added as artificial background.
|
||||
if adding_rgb_background:
|
||||
dir_rgb_backgrounds = None # Directory of texture images for synthetic background
|
||||
adding_rgb_foreground = False # Whether texture images will be added as artificial foreground.
|
||||
if adding_rgb_foreground:
|
||||
dir_rgb_foregrounds = None # Directory of texture images for synthetic foreground
|
||||
if adding_rgb_background or adding_rgb_foreground:
|
||||
number_of_backgrounds_per_image = 1
|
||||
if task == 'cnn-rnn-ocr':
|
||||
image_inversion = False # Whether the binarized images will be inverted.
|
||||
textline_skewing_bin = False # Whether binarized textline images will be rotated.
|
||||
textline_left_in_depth_bin = False # Whether left side of binary textline image will be displayed in depth.
|
||||
textline_right_in_depth_bin = False # Whether right side of binary textline image will be displayed in depth.
|
||||
textline_up_in_depth_bin = False # Whether upper side of binary textline image will be displayed in depth.
|
||||
textline_down_in_depth_bin = False # Whether lower side of binary textline image will be displayed in depth.
|
||||
pepper_bin_aug = False # Whether pepper noise will be added to binary textline images.
|
||||
bin_deg = False # Whether a combination of degrading and binarization will be applied to the image.
|
||||
degrading = False # Whether images will be artificially degraded. Requires the "degrade_scales" setting.
|
||||
if degrading or binarization and task == 'cnn-rnn-ocr' and bin_deg:
|
||||
degrade_scales = None # List of quality factors for degradation.
|
||||
channels_shuffling = False # Re-arrange color channels.
|
||||
if channels_shuffling:
|
||||
shuffle_indexes = None # List of channels to switch between.
|
||||
rotation = False # Whether images will be rotated by 90 degrees.
|
||||
rotation_not_90 = False # Whether images will be rotated arbitrarily (skewed). Requires "thetha" setting.
|
||||
if rotation_not_90:
|
||||
thetha = None # List of rotation angles in degrees.
|
||||
if task == 'cnn-rnn-ocr':
|
||||
white_noise_strap = False # Whether white noise will be applied on some straps on the textline image.
|
||||
textline_skewing = False # Whether textline images will be skewed for augmentation.
|
||||
if textline_skewing or binarization and textline_skewing_bin:
|
||||
skewing_amplitudes = None # List of skewing angles in degrees like [5, 8]
|
||||
textline_left_in_depth = False # If true, left side of textline image will be displayed in depth.
|
||||
textline_right_in_depth = False # If true, right side of textline image will be displayed in depth.
|
||||
textline_up_in_depth = False # If true, upper side of textline image will be displayed in depth.
|
||||
textline_down_in_depth = False # If true, lower side of textline image will be displayed in depth.
|
||||
pepper_aug = False # Whether pepper noise will be added to textline images.
|
||||
if pepper_aug or binarization and pepper_bin_aug:
|
||||
pepper_indexes = None # List of pepper noise factors, e.g. [0.01, 0.005].
|
||||
color_padding_rotation = False # Whether images will be rotated with color padding. Requires "thetha_padd" setting.
|
||||
if color_padding_rotation:
|
||||
thetha_padd = None # List of angles (in degrees) used for rotation alongside padding.
|
||||
dir_train = None # Directory of training dataset with subdirectories having the names "images" and "labels".
|
||||
dir_eval = None # Directory of validation dataset with subdirectories having the names "images" and "labels".
|
||||
dir_output = None # Directory where the augmented training data and the model checkpoints will be saved.
|
||||
pretraining = False # Set to true to (down)load pretrained weights of ResNet50 encoder.
|
||||
save_interval = None # frequency for writing model checkpoints (positive integer for number of batches saved under "model_step_{batch:04d}", otherwise epoch saved under "model_{epoch:02d}")
|
||||
reload_weights = False # Set true to build new model from config, load weights from dir_of_start_model, save under dir_output and exit.
|
||||
continue_training = False # Whether to continue training an existing model.
|
||||
if continue_training:
|
||||
dir_of_start_model = '' # Directory of model checkpoint to load to continue training. (E.g. if you already trained for 3 epochs, set "dir_of_start_model=dir_output/model_03".)
|
||||
index_start = 0 # Epoch counter initial value to continue training. (E.g. if you already trained for 3 epochs, set "index_start=3" to continue naming checkpoints model_04, model_05 etc.)
|
||||
data_is_provided = False # Whether the preprocessed input data (subdirectories "images" and "labels" in both subdirectories "train" and "eval" of "dir_output") has already been generated (in the first epoch of a previous run).
|
||||
|
||||
@ex.main
|
||||
def run(_config,
|
||||
_log,
|
||||
task,
|
||||
pretraining,
|
||||
data_is_provided,
|
||||
dir_train,
|
||||
dir_eval,
|
||||
dir_output,
|
||||
n_classes,
|
||||
n_epochs,
|
||||
n_batch,
|
||||
input_height,
|
||||
input_width,
|
||||
weight_decay,
|
||||
learning_rate,
|
||||
continue_training,
|
||||
reload_weights,
|
||||
save_interval,
|
||||
augmentation,
|
||||
# dependent config keys need a default,
|
||||
# otherwise yields sacred.utils.ConfigAddedError
|
||||
## if rotation_not_90
|
||||
thetha=None,
|
||||
is_loss_soft_dice=False,
|
||||
weighted_loss=False,
|
||||
add_ncc_loss=None,
|
||||
## if continue_training
|
||||
index_start=0,
|
||||
dir_of_start_model=None,
|
||||
backbone_type=None,
|
||||
## if backbone_type=transformer
|
||||
transformer_projection_dim=None,
|
||||
transformer_mlp_head_units=None,
|
||||
transformer_layers=None,
|
||||
transformer_num_heads=None,
|
||||
transformer_cnn_first=None,
|
||||
transformer_patchsize_x=None,
|
||||
transformer_patchsize_y=None,
|
||||
transformer_num_patches_xy=None,
|
||||
## if task=classification
|
||||
f1_threshold_classification=None,
|
||||
classification_classes_name=None,
|
||||
## if task=cnn-rnn-ocr
|
||||
characters_txt_file=None,
|
||||
color_padding_rotation=False,
|
||||
thetha_padd=None,
|
||||
bin_deg=False,
|
||||
image_inversion=False,
|
||||
white_noise_strap=False,
|
||||
textline_skewing=False,
|
||||
textline_skewing_bin=False,
|
||||
textline_left_in_depth=False,
|
||||
textline_left_in_depth_bin=False,
|
||||
textline_right_in_depth=False,
|
||||
textline_right_in_depth_bin=False,
|
||||
textline_up_in_depth=False,
|
||||
textline_up_in_depth_bin=False,
|
||||
textline_down_in_depth=False,
|
||||
textline_down_in_depth_bin=False,
|
||||
pepper_aug=False,
|
||||
pepper_bin_aug=False,
|
||||
pepper_indexes=None,
|
||||
padd_colors=None,
|
||||
white_padds=None,
|
||||
skewing_amplitudes=None,
|
||||
max_len=None,
|
||||
):
|
||||
"""
|
||||
run configured experiment via sacred
|
||||
"""
|
||||
|
||||
if continue_training:
|
||||
assert n_epochs > index_start, "with continue_training, n_epochs must be greater than index_start"
|
||||
|
||||
if pretraining and not os.path.isfile(RESNET50_WEIGHTS_PATH):
|
||||
_log.info("downloading RESNET50 pretrained weights to %s", RESNET50_WEIGHTS_PATH)
|
||||
download_file(RESNET50_WEIGHTS_URL, RESNET50_WEIGHTS_PATH)
|
||||
|
||||
# set the gpu configuration
|
||||
configuration()
|
||||
|
||||
if task in ["segmentation", "enhancement", "binarization"]:
|
||||
dir_train_flowing = os.path.join(dir_output, 'train')
|
||||
dir_eval_flowing = os.path.join(dir_output, 'eval')
|
||||
|
||||
dir_flow_train_imgs = os.path.join(dir_train_flowing, 'images')
|
||||
dir_flow_train_labels = os.path.join(dir_train_flowing, 'labels')
|
||||
|
||||
dir_flow_eval_imgs = os.path.join(dir_eval_flowing, 'images')
|
||||
dir_flow_eval_labels = os.path.join(dir_eval_flowing, 'labels')
|
||||
|
||||
if weighted_loss:
|
||||
weights = np.zeros(n_classes)
|
||||
if data_is_provided:
|
||||
dirs = dir_flow_train_labels
|
||||
else:
|
||||
dirs = os.path.join(dir_train, "labels")
|
||||
for obj in os.listdir(dirs):
|
||||
label_file = os.path.join(dirs, + obj)
|
||||
try:
|
||||
label_obj = cv2.imread(label_file)
|
||||
label_obj_one_hot = get_one_hot(label_obj, label_obj.shape[0], label_obj.shape[1], n_classes)
|
||||
weights += (label_obj_one_hot.sum(axis=0)).sum(axis=0)
|
||||
except Exception:
|
||||
_log.exception("error reading data file '%s'", label_file)
|
||||
|
||||
weights = 1.00 / weights
|
||||
weights = weights / float(np.sum(weights))
|
||||
weights = weights / float(np.min(weights))
|
||||
weights = weights / float(np.sum(weights))
|
||||
|
||||
if task == "enhancement":
|
||||
assert not is_loss_soft_dice, "for enhancement, soft_dice loss does not apply"
|
||||
assert not weighted_loss, "for enhancement, weighted loss does not apply"
|
||||
if continue_training:
|
||||
custom_objects = dict()
|
||||
if is_loss_soft_dice:
|
||||
custom_objects.update(soft_dice_loss=soft_dice_loss)
|
||||
elif weighted_loss:
|
||||
custom_objects.update(loss=weighted_categorical_crossentropy(weights))
|
||||
if backbone_type == 'transformer':
|
||||
custom_objects.update(PatchEncoder=PatchEncoder,
|
||||
Patches=Patches)
|
||||
model = load_model(dir_of_start_model, compile=False,
|
||||
custom_objects=custom_objects)
|
||||
else:
|
||||
index_start = 0
|
||||
if backbone_type == 'nontransformer':
|
||||
model = resnet50_unet(n_classes,
|
||||
input_height,
|
||||
input_width,
|
||||
task,
|
||||
weight_decay,
|
||||
pretraining)
|
||||
else:
|
||||
num_patches_x = transformer_num_patches_xy[0]
|
||||
num_patches_y = transformer_num_patches_xy[1]
|
||||
num_patches = num_patches_x * num_patches_y
|
||||
|
||||
if transformer_cnn_first:
|
||||
model_builder = vit_resnet50_unet
|
||||
multiple = 32
|
||||
else:
|
||||
model_builder = vit_resnet50_unet_transformer_before_cnn
|
||||
multiple = 1
|
||||
|
||||
assert input_height == (
|
||||
num_patches_y * transformer_patchsize_y * multiple), (
|
||||
"transformer_patchsize_y or transformer_num_patches_xy height value error: "
|
||||
"input_height should be equal to "
|
||||
"(transformer_num_patches_xy height value * transformer_patchsize_y * %d)" % multiple)
|
||||
assert input_width == (
|
||||
num_patches_x * transformer_patchsize_x * multiple), (
|
||||
"transformer_patchsize_x or transformer_num_patches_xy width value error: "
|
||||
"input_width should be equal to "
|
||||
"(transformer_num_patches_xy width value * transformer_patchsize_x * %d)" % multiple)
|
||||
assert 0 == (transformer_projection_dim %
|
||||
(transformer_patchsize_y * transformer_patchsize_x)), (
|
||||
"transformer_projection_dim error: "
|
||||
"The remainder when parameter transformer_projection_dim is divided by "
|
||||
"(transformer_patchsize_y*transformer_patchsize_x) should be zero")
|
||||
|
||||
model_builder = create_captured_function(model_builder)
|
||||
model_builder.config = _config
|
||||
model_builder.logger = _log
|
||||
model = model_builder(num_patches)
|
||||
|
||||
assert model is not None
|
||||
#if you want to see the model structure just uncomment model summary.
|
||||
#model.summary()
|
||||
|
||||
metrics = ['categorical_accuracy']
|
||||
if task in ["segmentation", "binarization"]:
|
||||
if is_loss_soft_dice:
|
||||
loss = soft_dice_loss
|
||||
elif weighted_loss:
|
||||
loss = weighted_categorical_crossentropy(weights)
|
||||
else:
|
||||
loss = get_metric('categorical_crossentropy')
|
||||
if add_ncc_loss:
|
||||
loss = metrics_superposition(loss, connected_components_loss(n_classes - 1),
|
||||
weights=[1 - add_ncc_loss, add_ncc_loss])
|
||||
metrics.append(connected_components_loss(n_classes - 1))
|
||||
metrics.append(MeanIoU(n_classes,
|
||||
name='iou',
|
||||
ignore_class=0,
|
||||
sparse_y_true=False,
|
||||
sparse_y_pred=False))
|
||||
metrics.append(ConfusionMatrix(n_classes))
|
||||
else: # task == "enhancement"
|
||||
loss = 'mean_squared_error'
|
||||
model.compile(loss=loss,
|
||||
#jit_compile=True,
|
||||
optimizer=Adam(learning_rate=learning_rate),
|
||||
metrics=metrics)
|
||||
|
||||
if reload_weights:
|
||||
model.load_weights(dir_of_start_model).assert_existing_objects_matched().expect_partial()
|
||||
dir_save = os.path.join(dir_output, os.path.basename(os.path.normpath(dir_of_start_model)))
|
||||
model.save(dir_save, include_optimizer=False)
|
||||
with open(os.path.join(dir_save, "config.json"), "w") as fp:
|
||||
json.dump(_config, fp) # encode dict into JSON
|
||||
_log.info("reloaded model from %s to %s", dir_of_start_model, dir_save)
|
||||
return
|
||||
|
||||
if not data_is_provided:
|
||||
# first create a directory in output for both training and evaluations
|
||||
# in order to flow data from these directories.
|
||||
if os.path.isdir(dir_train_flowing):
|
||||
os.system('rm -rf ' + dir_train_flowing)
|
||||
os.makedirs(dir_train_flowing)
|
||||
|
||||
if os.path.isdir(dir_eval_flowing):
|
||||
os.system('rm -rf ' + dir_eval_flowing)
|
||||
os.makedirs(dir_eval_flowing)
|
||||
|
||||
os.mkdir(dir_flow_train_imgs)
|
||||
os.mkdir(dir_flow_train_labels)
|
||||
|
||||
os.mkdir(dir_flow_eval_imgs)
|
||||
os.mkdir(dir_flow_eval_labels)
|
||||
|
||||
# writing patches into a sub-folder in order to be flowed from directory.
|
||||
def gen(dir_img, dir_lab, dir_flow_imgs, dir_flow_labs, augmentation=True):
|
||||
indexer = 0
|
||||
for img, lab in tqdm(preprocess_imgs(_config,
|
||||
dir_img,
|
||||
dir_lab,
|
||||
augmentation=augmentation),
|
||||
desc="data_is_provided"):
|
||||
fname = 'img_%d.png' % indexer
|
||||
cv2.imwrite(os.path.join(dir_flow_imgs, fname), img)
|
||||
cv2.imwrite(os.path.join(dir_flow_labs, fname), lab)
|
||||
indexer += 1
|
||||
gen(*get_dirs_or_files(dir_train),
|
||||
dir_flow_train_imgs,
|
||||
dir_flow_train_labels)
|
||||
gen(*get_dirs_or_files(dir_eval),
|
||||
dir_flow_eval_imgs,
|
||||
dir_flow_eval_labels,
|
||||
augmentation=False)
|
||||
|
||||
def _to_cv2float(img):
|
||||
# rgb→bgr and uint8→float, as expected by Eynollah models
|
||||
return tf.cast(tf.reverse(img, [-1]), tf.float32) / 255
|
||||
def _to_intrgb(img):
|
||||
# bgr→rgb and float→uint8 for plotting
|
||||
return tf.reverse(tf.cast(img * 255, tf.uint8), [-1])
|
||||
def _to_categorical(seg):
|
||||
seg = tf.cast(seg * 255, tf.int8)
|
||||
# gt_gen_utils/pagexml2label uses peculiar pseudo-RGB/index colors
|
||||
#seg = tf.image.rgb_to_grayscale(seg)
|
||||
seg = tf.gather(seg, [0], axis=-1)
|
||||
seg = tf.squeeze(seg, axis=-1)
|
||||
return one_hot(seg, n_classes)
|
||||
def get_dataset(dir_imgs, dir_labs, shuffle=None):
|
||||
gen_kwargs = dict(labels=None,
|
||||
label_mode=None,
|
||||
batch_size=None, # batch after zip below
|
||||
image_size=(input_height, input_width),
|
||||
color_mode='rgb',
|
||||
shuffle=shuffle is not None,
|
||||
seed=shuffle,
|
||||
interpolation='nearest',
|
||||
crop_to_aspect_ratio=False,
|
||||
# Keras 3 only...
|
||||
#pad_to_aspect_ratio=False,
|
||||
#data_format='channel_last',
|
||||
#verbose=False,
|
||||
)
|
||||
img_gen = image_dataset_from_directory(dir_imgs, **gen_kwargs)
|
||||
lab_gen = image_dataset_from_directory(dir_labs, **gen_kwargs)
|
||||
img_gen = img_gen.map(_to_cv2float, num_parallel_calls=tf.data.AUTOTUNE)
|
||||
lab_gen = lab_gen.map(_to_cv2float, num_parallel_calls=tf.data.AUTOTUNE)
|
||||
if task in ["segmentation", "binarization"]:
|
||||
lab_gen = lab_gen.map(_to_categorical, num_parallel_calls=tf.data.AUTOTUNE)
|
||||
ds = tf.data.Dataset.zip(img_gen, lab_gen)
|
||||
return ds.batch(n_batch, drop_remainder=True, num_parallel_calls=tf.data.AUTOTUNE)
|
||||
train_gen = get_dataset(dir_flow_train_imgs, dir_flow_train_labels, shuffle=np.random.randint(1e6))
|
||||
valdn_gen = get_dataset(dir_flow_eval_imgs, dir_flow_eval_labels)
|
||||
train_steps = len(os.listdir(dir_flow_train_imgs)) // n_batch
|
||||
valdn_steps = len(os.listdir(dir_flow_eval_imgs)) // n_batch
|
||||
_log.info("training on %d batches in %d epochs", train_steps, n_epochs)
|
||||
_log.info("validating on %d batches", valdn_steps)
|
||||
|
||||
callbacks = [TensorBoardPlotter((max(1, train_steps * n_batch // 1000),
|
||||
max(1, valdn_steps * n_batch // 100)),
|
||||
os.path.join(dir_output, 'logs'),
|
||||
profile_batch=(10, 20)),
|
||||
SaveWeightsAfterSteps(0, dir_output, _config),
|
||||
]
|
||||
if save_interval:
|
||||
callbacks.append(SaveWeightsAfterSteps(save_interval, dir_output, _config))
|
||||
train_gen = train_gen.shuffle(train_steps // 1000, reshuffle_each_iteration=True)
|
||||
valdn_gen = valdn_gen.shuffle(valdn_steps // 10, reshuffle_each_iteration=False)
|
||||
# from matplotlib import pyplot as plt
|
||||
# from tensorflow_addons.image import connected_components
|
||||
# def plot(x, ytrue):
|
||||
# ypred = model.call(x)
|
||||
# gt = plot_layout_tf(x, ytrue)
|
||||
# dt = plot_layout_tf(x, ypred)
|
||||
# segtrue = tf.math.argmax(ytrue, axis=-1)
|
||||
# segpred = tf.math.argmax(ypred, axis=-1)
|
||||
# cctrue = connected_components(segtrue)
|
||||
# ccpred = connected_components(segpred)
|
||||
# cc = connected_components_loss(n_classes-1)(ytrue, ypred)
|
||||
# sd = soft_dice_loss(ytrue, ypred)
|
||||
# return gt, dt, cctrue, ccpred, cc, sd
|
||||
# for gt, dt, gtcc, dtcc, cc, sd in train_gen.take(15).rebatch(1).map(plot).as_numpy_iterator():
|
||||
# plt.subplot(2, 2, 1)
|
||||
# plt.imshow(np.squeeze(gt))
|
||||
# plt.title('GT')
|
||||
# plt.subplot(2, 2, 3)
|
||||
# plt.imshow(np.squeeze(gtcc))
|
||||
# plt.title('GT CC')
|
||||
# plt.subplot(2, 2, 4)
|
||||
# plt.imshow(np.squeeze(dtcc))
|
||||
# plt.title('prediction CC')
|
||||
# plt.subplot(2, 2, 2)
|
||||
# plt.imshow(np.squeeze(dt))
|
||||
# plt.title(f'prediction (nCC={cc} soft dice={sd:.3f})')
|
||||
# plt.show()
|
||||
model.fit(
|
||||
train_gen.prefetch(tf.data.AUTOTUNE),
|
||||
steps_per_epoch=train_steps,
|
||||
validation_data=valdn_gen.prefetch(tf.data.AUTOTUNE),
|
||||
validation_steps=valdn_steps,
|
||||
verbose=1,
|
||||
epochs=n_epochs,
|
||||
callbacks=callbacks,
|
||||
initial_epoch=index_start)
|
||||
|
||||
elif task=="cnn-rnn-ocr":
|
||||
|
||||
with open(characters_txt_file, 'r') as char_txt_f:
|
||||
characters = json.load(char_txt_f)
|
||||
padding_token = len(characters) + 5
|
||||
# Mapping characters to integers.
|
||||
char_to_num = StringLookup(vocabulary=list(characters), mask_token=None)
|
||||
n_classes = len(char_to_num.get_vocabulary()) + 2
|
||||
|
||||
if continue_training:
|
||||
model = load_model(dir_of_start_model)
|
||||
else:
|
||||
index_start = 0
|
||||
model = cnn_rnn_ocr_model(image_height=input_height,
|
||||
image_width=input_width,
|
||||
n_classes=n_classes,
|
||||
max_seq=max_len)
|
||||
#initial_learning_rate = 1e-4
|
||||
#decay_steps = int (n_epochs * ( len_dataset / n_batch ))
|
||||
#alpha = 0.01
|
||||
#lr_schedule = 1e-4
|
||||
#tf.keras.optimizers.schedules.CosineDecay(initial_learning_rate, decay_steps, alpha)
|
||||
opt = Adam(learning_rate=learning_rate)
|
||||
model.compile(optimizer=opt) # rs: loss seems to be (ctc_batch_cost) in last layer
|
||||
|
||||
#print(model.summary())
|
||||
|
||||
if reload_weights:
|
||||
model.load_weights(dir_of_start_model).assert_existing_objects_matched().expect_partial()
|
||||
dir_save = os.path.join(dir_output, os.path.basename(os.path.normpath(dir_of_start_model)))
|
||||
model.save(dir_save, include_optimizer=False)
|
||||
with open(os.path.join(dir_save, "config.json"), "w") as fp:
|
||||
json.dump(_config, fp) # encode dict into JSON
|
||||
_log.info("reloaded model from %s to %s", dir_of_start_model, dir_save)
|
||||
return
|
||||
|
||||
# todo: use Dataset.map() on Dataset.list_files()
|
||||
def get_dataset(dir_img, dir_lab):
|
||||
def gen():
|
||||
return preprocess_imgs(_config,
|
||||
dir_img,
|
||||
dir_lab,
|
||||
# extra+overrides
|
||||
char_to_num=char_to_num,
|
||||
padding_token=padding_token
|
||||
)
|
||||
return (tf.data.Dataset.from_generator(gen, (tf.float32, tf.int64))
|
||||
.padded_batch(n_batch,
|
||||
padded_shapes=([input_height, input_width, 3], [None]),
|
||||
padding_values=(None, tf.constant(padding_token, dtype=tf.int64)),
|
||||
drop_remainder=True,
|
||||
#num_parallel_calls=tf.data.AUTOTUNE,
|
||||
)
|
||||
.map(lambda x, y: {"image": x, "label": y})
|
||||
.prefetch(tf.data.AUTOTUNE)
|
||||
)
|
||||
train_ds = get_dataset(*get_dirs_or_files(dir_train))
|
||||
valdn_ds = get_dataset(*get_dirs_or_files(dir_eval))
|
||||
|
||||
callbacks = [TensorBoard(os.path.join(dir_output, 'logs'), write_graph=False),
|
||||
EarlyStopping(verbose=1, patience=3, restore_best_weights=False, start_from_epoch=3),
|
||||
SaveWeightsAfterSteps(0, dir_output, _config)]
|
||||
if save_interval:
|
||||
callbacks.append(SaveWeightsAfterSteps(save_interval, dir_output, _config))
|
||||
model.fit(
|
||||
train_ds.shuffle(200),
|
||||
validation_data=valdn_ds,
|
||||
verbose=1,
|
||||
epochs=n_epochs,
|
||||
callbacks=callbacks,
|
||||
initial_epoch=index_start)
|
||||
|
||||
elif task=='classification':
|
||||
if continue_training:
|
||||
model = load_model(dir_of_start_model, compile=False)
|
||||
else:
|
||||
index_start = 0
|
||||
model = resnet50_classifier(n_classes,
|
||||
input_height,
|
||||
input_width,
|
||||
weight_decay,
|
||||
pretraining)
|
||||
|
||||
model.compile(loss='categorical_crossentropy',
|
||||
optimizer=Adam(learning_rate=0.001), # rs: why not learning_rate?
|
||||
metrics=['accuracy', F1Score(average='macro', name='f1')])
|
||||
|
||||
if reload_weights:
|
||||
model.load_weights(dir_of_start_model).assert_existing_objects_matched().expect_partial()
|
||||
dir_save = os.path.join(dir_output, os.path.basename(os.path.normpath(dir_of_start_model)))
|
||||
model.save(dir_save, include_optimizer=False)
|
||||
with open(os.path.join(dir_save, "config.json"), "w") as fp:
|
||||
json.dump(_config, fp) # encode dict into JSON
|
||||
_log.info("reloaded model from %s to %s", dir_of_start_model, dir_save)
|
||||
return
|
||||
|
||||
list_classes = list(classification_classes_name.values())
|
||||
data_args = dict(label_mode="categorical",
|
||||
class_names=list_classes,
|
||||
batch_size=n_batch,
|
||||
image_size=(input_height, input_width),
|
||||
interpolation="nearest")
|
||||
trainXY = image_dataset_from_directory(dir_train, shuffle=True, **data_args)
|
||||
testXY = image_dataset_from_directory(dir_eval, shuffle=False, **data_args)
|
||||
callbacks = [TensorBoard(os.path.join(dir_output, 'logs'), write_graph=False),
|
||||
SaveWeightsAfterSteps(0, dir_output, _config,
|
||||
monitor='val_f1',
|
||||
#save_best_only=True, # we need all for ensembling
|
||||
mode='max')]
|
||||
|
||||
history = model.fit(trainXY,
|
||||
#class_weight=weights)
|
||||
validation_data=testXY,
|
||||
verbose=1,
|
||||
epochs=n_epochs,
|
||||
callbacks=callbacks,
|
||||
initial_epoch=index_start)
|
||||
|
||||
usable_checkpoints = np.flatnonzero(np.array(history.history['val_f1']) >
|
||||
f1_threshold_classification)
|
||||
if len(usable_checkpoints) >= 1:
|
||||
_log.info("averaging over usable checkpoints: %s", str(usable_checkpoints))
|
||||
usable_checkpoints = [os.path.join(dir_output, 'model_{epoch:02d}'.format(epoch=epoch + 1))
|
||||
for epoch in usable_checkpoints]
|
||||
ens_path = os.path.join(dir_output, 'model_ens_avg')
|
||||
run_ensembling(usable_checkpoints, ens_path)
|
||||
_log.info("ensemble model saved under '%s'", ens_path)
|
||||
|
||||
elif task=='reading_order':
|
||||
if continue_training:
|
||||
model = load_model(dir_of_start_model, compile=False)
|
||||
else:
|
||||
index_start = 0
|
||||
model = machine_based_reading_order_model(n_classes,
|
||||
input_height,
|
||||
input_width,
|
||||
weight_decay,
|
||||
pretraining)
|
||||
|
||||
#f1score_tot = [0]
|
||||
model.compile(loss="binary_crossentropy",
|
||||
#optimizer=SGD(learning_rate=0.01, momentum=0.9),
|
||||
optimizer=Adam(learning_rate=0.0001), # rs: why not learning_rate?
|
||||
metrics=['accuracy'])
|
||||
|
||||
if reload_weights:
|
||||
model.load_weights(dir_of_start_model).assert_existing_objects_matched().expect_partial()
|
||||
dir_save = os.path.join(dir_output, os.path.basename(os.path.normpath(dir_of_start_model)))
|
||||
model.save(dir_save, include_optimizer=False)
|
||||
with open(os.path.join(dir_save, "config.json"), "w") as fp:
|
||||
json.dump(_config, fp) # encode dict into JSON
|
||||
_log.info("reloaded model from %s to %s", dir_of_start_model, dir_save)
|
||||
return
|
||||
|
||||
dir_flow_train_imgs = os.path.join(dir_train, 'images')
|
||||
dir_flow_train_labels = os.path.join(dir_train, 'labels')
|
||||
|
||||
classes = os.listdir(dir_flow_train_labels)
|
||||
if augmentation:
|
||||
num_rows = len(classes)*(len(thetha) + 1)
|
||||
else:
|
||||
num_rows = len(classes)
|
||||
#ls_test = os.listdir(dir_flow_train_labels)
|
||||
|
||||
callbacks = [TensorBoard(os.path.join(dir_output, 'logs'), write_graph=False),
|
||||
SaveWeightsAfterSteps(0, dir_output, _config)]
|
||||
if save_interval:
|
||||
callbacks.append(SaveWeightsAfterSteps(save_interval, dir_output, _config))
|
||||
|
||||
trainXY = generate_arrays_from_folder_reading_order(
|
||||
dir_flow_train_labels, dir_flow_train_imgs,
|
||||
n_batch, input_height, input_width, n_classes,
|
||||
thetha, augmentation)
|
||||
|
||||
history = model.fit(trainXY,
|
||||
steps_per_epoch=num_rows / n_batch,
|
||||
verbose=1,
|
||||
epochs=n_epochs,
|
||||
callbacks=callbacks,
|
||||
initial_epoch=index_start)
|
||||
'''
|
||||
if f1score>f1score_tot[0]:
|
||||
f1score_tot[0] = f1score
|
||||
model_dir = os.path.join(dir_out,'model_best')
|
||||
model.save(model_dir)
|
||||
'''
|
||||
File diff suppressed because it is too large
Load diff
|
|
@ -1,66 +0,0 @@
|
|||
import os
|
||||
from warnings import catch_warnings, simplefilter
|
||||
|
||||
import click
|
||||
import numpy as np
|
||||
|
||||
os.environ['TF_USE_LEGACY_KERAS'] = '1' # avoid Keras 3 after TF 2.15
|
||||
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
|
||||
|
||||
from ocrd_utils import tf_disable_interactive_logs
|
||||
tf_disable_interactive_logs()
|
||||
import tensorflow as tf
|
||||
from tensorflow.keras.models import load_model
|
||||
|
||||
from ..patch_encoder import (
|
||||
PatchEncoder,
|
||||
Patches,
|
||||
)
|
||||
|
||||
def run_ensembling(model_dirs, out_dir):
|
||||
all_weights = []
|
||||
|
||||
for model_dir in model_dirs:
|
||||
assert os.path.isdir(model_dir), model_dir
|
||||
model = load_model(model_dir, compile=False,
|
||||
custom_objects=dict(PatchEncoder=PatchEncoder,
|
||||
Patches=Patches))
|
||||
all_weights.append(model.get_weights())
|
||||
|
||||
new_weights = []
|
||||
for layer_weights in zip(*all_weights):
|
||||
layer_weights = np.array([np.array(weights).mean(axis=0)
|
||||
for weights in zip(*layer_weights)])
|
||||
new_weights.append(layer_weights)
|
||||
|
||||
#model = tf.keras.models.clone_model(model)
|
||||
model.set_weights(new_weights)
|
||||
|
||||
model.save(out_dir)
|
||||
os.system('cp ' + os.path.join(model_dirs[0], "config.json ") + out_dir + "/")
|
||||
|
||||
@click.command()
|
||||
@click.option(
|
||||
"--in",
|
||||
"-i",
|
||||
help="input directory of checkpoint models to be read",
|
||||
multiple=True,
|
||||
required=True,
|
||||
type=click.Path(exists=True, file_okay=False),
|
||||
)
|
||||
@click.option(
|
||||
"--out",
|
||||
"-o",
|
||||
help="output directory where ensembled model will be written.",
|
||||
required=True,
|
||||
type=click.Path(exists=False, file_okay=False),
|
||||
)
|
||||
def ensemble_cli(in_, out):
|
||||
"""
|
||||
mix multiple model weights
|
||||
|
||||
Load a sequence of models and mix them into a single ensemble model
|
||||
by averaging their weights. Write the resulting model.
|
||||
"""
|
||||
run_ensembling(in_, out)
|
||||
|
||||
File diff suppressed because it is too large
Load diff
|
|
@ -1,546 +0,0 @@
|
|||
from typing import Sequence, Union
|
||||
from numbers import Number
|
||||
from functools import partial
|
||||
import itertools
|
||||
|
||||
import cv2
|
||||
import numpy as np
|
||||
from scipy.sparse.csgraph import minimum_spanning_tree
|
||||
from shapely.geometry import Polygon, LineString
|
||||
from shapely.geometry.polygon import orient
|
||||
from shapely import set_precision, affinity
|
||||
from shapely.ops import unary_union, nearest_points
|
||||
|
||||
from .rotate import rotate_image, rotation_image_new
|
||||
|
||||
def contours_in_same_horizon(cy_main_hor):
|
||||
"""
|
||||
Takes an array of y coords, identifies all pairs among them
|
||||
which are close to each other, and returns all such pairs
|
||||
by index into the array.
|
||||
"""
|
||||
sort = np.argsort(cy_main_hor)
|
||||
same = np.diff(cy_main_hor[sort]) <= 20
|
||||
# groups = np.split(sort, np.arange(len(cy_main_hor) - 1)[~same] + 1)
|
||||
same = np.flatnonzero(same)
|
||||
return np.stack((sort[:-1][same], sort[1:][same])).T
|
||||
|
||||
def find_contours_mean_y_diff(contours_main):
|
||||
M_main = [cv2.moments(contours_main[j]) for j in range(len(contours_main))]
|
||||
cy_main = [(M_main[j]["m01"] / (M_main[j]["m00"] + 1e-32)) for j in range(len(M_main))]
|
||||
return np.mean(np.diff(np.sort(np.array(cy_main))))
|
||||
|
||||
def get_text_region_boxes_by_given_contours(contours):
|
||||
return [cv2.boundingRect(contour)
|
||||
for contour in contours]
|
||||
|
||||
def filter_contours_area_of_image(image, contours, hierarchy, max_area=1.0, min_area=0.0, dilate=0):
|
||||
found_polygons_early = []
|
||||
for jv, contour in enumerate(contours):
|
||||
if len(contour) < 3: # A polygon cannot have less than 3 points
|
||||
continue
|
||||
|
||||
polygon = contour2polygon(contour, dilate=dilate)
|
||||
area = polygon.area
|
||||
if (area >= min_area * np.prod(image.shape[:2]) and
|
||||
area <= max_area * np.prod(image.shape[:2]) and
|
||||
hierarchy[0][jv][3] == -1):
|
||||
found_polygons_early.append(polygon2contour(polygon))
|
||||
return found_polygons_early
|
||||
|
||||
def filter_contours_area_of_image_tables(image, contours, hierarchy, max_area=1.0, min_area=0.0, dilate=0):
|
||||
found_polygons_early = []
|
||||
for jv, contour in enumerate(contours):
|
||||
if len(contour) < 3: # A polygon cannot have less than 3 points
|
||||
continue
|
||||
|
||||
polygon = contour2polygon(contour, dilate=dilate)
|
||||
# area = cv2.contourArea(contour)
|
||||
area = polygon.area
|
||||
##print(np.prod(thresh.shape[:2]))
|
||||
# Check that polygon has area greater than minimal area
|
||||
# print(hierarchy[0][jv][3],hierarchy )
|
||||
if (area >= min_area * image.size and
|
||||
area <= max_area * image.size and
|
||||
# hierarchy[0][jv][3]==-1
|
||||
True):
|
||||
# print(contour[0][0][1])
|
||||
found_polygons_early.append(polygon2contour(polygon))
|
||||
return found_polygons_early
|
||||
|
||||
def find_center_of_contours(contours):
|
||||
moments = [cv2.moments(contour) for contour in contours]
|
||||
cx = [feat["m10"] / (feat["m00"] + 1e-32)
|
||||
for feat in moments]
|
||||
cy = [feat["m01"] / (feat["m00"] + 1e-32)
|
||||
for feat in moments]
|
||||
return cx, cy
|
||||
|
||||
def find_new_features_of_contours(contours):
|
||||
# areas = np.array([cv2.contourArea(contour) for contour in contours])
|
||||
cx, cy = find_center_of_contours(contours)
|
||||
slice_x = np.index_exp[:, 0, 0]
|
||||
slice_y = np.index_exp[:, 0, 1]
|
||||
if any(contour.ndim < 3 for contour in contours):
|
||||
slice_x = np.index_exp[:, 0]
|
||||
slice_y = np.index_exp[:, 1]
|
||||
x_min = np.array([np.min(contour[slice_x]) for contour in contours])
|
||||
x_max = np.array([np.max(contour[slice_x]) for contour in contours])
|
||||
y_min = np.array([np.min(contour[slice_y]) for contour in contours])
|
||||
y_max = np.array([np.max(contour[slice_y]) for contour in contours])
|
||||
# dis_x=np.abs(x_max-x_min)
|
||||
y_corr_x_min = np.array([contour[np.argmin(contour[slice_x])][slice_y[1:]]
|
||||
for contour in contours])
|
||||
|
||||
return cx, cy, x_min, x_max, y_min, y_max, y_corr_x_min
|
||||
|
||||
def find_features_of_contours(contours):
|
||||
y_min = np.array([np.min(contour[:,0,1]) for contour in contours])
|
||||
y_max = np.array([np.max(contour[:,0,1]) for contour in contours])
|
||||
|
||||
return y_min, y_max
|
||||
|
||||
def return_parent_contours(contours, hierarchy):
|
||||
contours_parent = [contours[i]
|
||||
for i in range(len(contours))
|
||||
if hierarchy[0][i][3] == -1]
|
||||
return contours_parent
|
||||
|
||||
def return_contours_of_interested_region(region_pre_p, label, min_area=0.0002, dilate=0):
|
||||
if region_pre_p.ndim == 3:
|
||||
mask = (region_pre_p[:, :, 0] == label).astype(np.uint8)
|
||||
else:
|
||||
mask = (region_pre_p[:, :] == label).astype(np.uint8)
|
||||
|
||||
contours_imgs, hierarchy = cv2.findContours(mask, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
|
||||
contours_imgs = return_parent_contours(contours_imgs, hierarchy)
|
||||
contours_imgs = filter_contours_area_of_image_tables(mask, contours_imgs, hierarchy,
|
||||
max_area=1,
|
||||
min_area=min_area,
|
||||
dilate=dilate)
|
||||
return contours_imgs
|
||||
|
||||
def do_work_of_contours_in_image(contour, index_r_con, img, slope_first):
|
||||
img_copy = np.zeros(img.shape[:2], dtype=np.uint8)
|
||||
img_copy = cv2.fillPoly(img_copy, pts=[contour], color=1)
|
||||
|
||||
img_copy = rotation_image_new(img_copy, -slope_first)
|
||||
_, thresh = cv2.threshold(img_copy, 0, 255, 0)
|
||||
|
||||
cont_int, _ = cv2.findContours(thresh, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
|
||||
|
||||
cont_int[0][:, 0, 0] = cont_int[0][:, 0, 0] + np.abs(img_copy.shape[1] - img.shape[1])
|
||||
cont_int[0][:, 0, 1] = cont_int[0][:, 0, 1] + np.abs(img_copy.shape[0] - img.shape[0])
|
||||
|
||||
return cont_int[0], index_r_con
|
||||
|
||||
def get_textregion_contours_in_org_image_multi(cnts, img, slope_first, map=map):
|
||||
if not len(cnts):
|
||||
return [], []
|
||||
results = map(partial(do_work_of_contours_in_image,
|
||||
img=img,
|
||||
slope_first=slope_first,
|
||||
),
|
||||
cnts, range(len(cnts)))
|
||||
return tuple(zip(*results))
|
||||
|
||||
def get_textregion_contours_in_org_image(cnts, img, slope_first):
|
||||
cnts_org = []
|
||||
# print(cnts,'cnts')
|
||||
for i in range(len(cnts)):
|
||||
img_copy = np.zeros(img.shape[:2], dtype=np.uint8)
|
||||
img_copy = cv2.fillPoly(img_copy, pts=[cnts[i]], color=1)
|
||||
|
||||
# plt.imshow(img_copy)
|
||||
# plt.show()
|
||||
|
||||
# print(img.shape,'img')
|
||||
img_copy = rotation_image_new(img_copy, -slope_first)
|
||||
##print(img_copy.shape,'img_copy')
|
||||
# plt.imshow(img_copy)
|
||||
# plt.show()
|
||||
|
||||
_, thresh = cv2.threshold(img_copy, 0, 255, 0)
|
||||
|
||||
cont_int, _ = cv2.findContours(thresh, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
|
||||
cont_int[0][:, 0, 0] = cont_int[0][:, 0, 0] + np.abs(img_copy.shape[1] - img.shape[1])
|
||||
cont_int[0][:, 0, 1] = cont_int[0][:, 0, 1] + np.abs(img_copy.shape[0] - img.shape[0])
|
||||
# print(np.shape(cont_int[0]))
|
||||
cnts_org.append(cont_int[0])
|
||||
|
||||
return cnts_org
|
||||
|
||||
def get_textregion_confidences_old(cnts, img, slope_first):
|
||||
zoom = 3
|
||||
img = cv2.resize(img, (img.shape[1] // zoom,
|
||||
img.shape[0] // zoom),
|
||||
interpolation=cv2.INTER_NEAREST)
|
||||
cnts_org = []
|
||||
for cnt in cnts:
|
||||
img_copy = np.zeros(img.shape[:2], dtype=np.uint8)
|
||||
img_copy = cv2.fillPoly(img_copy, pts=[cnt // zoom], color=1)
|
||||
|
||||
img_copy = rotation_image_new(img_copy, -slope_first).astype(np.uint8)
|
||||
_, thresh = cv2.threshold(img_copy, 0, 255, 0)
|
||||
|
||||
cont_int, _ = cv2.findContours(thresh, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
|
||||
cont_int[0][:, 0, 0] = cont_int[0][:, 0, 0] + np.abs(img_copy.shape[1] - img.shape[1])
|
||||
cont_int[0][:, 0, 1] = cont_int[0][:, 0, 1] + np.abs(img_copy.shape[0] - img.shape[0])
|
||||
cnts_org.append(cont_int[0] * zoom)
|
||||
|
||||
return cnts_org
|
||||
|
||||
def do_back_rotation_and_get_cnt_back(contour_par, index_r_con, img, slope_first, confidence_matrix):
|
||||
img_copy = np.zeros(img.shape[:2], dtype=np.uint8)
|
||||
img_copy = cv2.fillPoly(img_copy, pts=[contour_par], color=1)
|
||||
confidence_matrix_mapped_with_contour = confidence_matrix * img_copy
|
||||
confidence_contour = np.sum(confidence_matrix_mapped_with_contour) / float(np.sum(img_copy))
|
||||
|
||||
img_copy = rotation_image_new(img_copy, -slope_first).astype(np.uint8)
|
||||
_, thresh = cv2.threshold(img_copy, 0, 255, 0)
|
||||
|
||||
cont_int, _ = cv2.findContours(thresh, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
|
||||
if len(cont_int)==0:
|
||||
cont_int = [contour_par]
|
||||
confidence_contour = 0
|
||||
else:
|
||||
cont_int[0][:, 0, 0] = cont_int[0][:, 0, 0] + np.abs(img_copy.shape[1] - img.shape[1])
|
||||
cont_int[0][:, 0, 1] = cont_int[0][:, 0, 1] + np.abs(img_copy.shape[0] - img.shape[0])
|
||||
return cont_int[0], index_r_con, confidence_contour
|
||||
|
||||
def get_region_confidences(cnts, confidence_matrix):
|
||||
if not len(cnts):
|
||||
return []
|
||||
|
||||
height, width = confidence_matrix.shape
|
||||
confidence_matrix = cv2.resize(confidence_matrix,
|
||||
(width // 6, height // 6),
|
||||
interpolation=cv2.INTER_NEAREST)
|
||||
confs = []
|
||||
for cnt in cnts:
|
||||
cnt_mask = np.zeros_like(confidence_matrix)
|
||||
cnt_mask = cv2.fillPoly(cnt_mask, pts=[cnt // 6], color=1.0)
|
||||
confs.append(np.sum(confidence_matrix * cnt_mask) / np.sum(cnt_mask))
|
||||
return confs
|
||||
|
||||
def return_contours_of_interested_textline(region_pre_p, label, min_area=0.0):
|
||||
cnts_images = (region_pre_p == label).astype(np.uint8)
|
||||
contours_imgs, hierarchy = cv2.findContours(cnts_images, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
|
||||
contours_imgs = return_parent_contours(contours_imgs, hierarchy)
|
||||
contours_imgs = filter_contours_area_of_image_tables(
|
||||
cnts_images, contours_imgs, hierarchy, max_area=1, min_area=min_area)
|
||||
return contours_imgs
|
||||
|
||||
def return_contours_of_image(image):
|
||||
if len(image.shape) == 2:
|
||||
image = image.astype(np.uint8)
|
||||
imgray = image
|
||||
else:
|
||||
image = image.astype(np.uint8)
|
||||
imgray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
|
||||
_, thresh = cv2.threshold(imgray, 0, 255, 0)
|
||||
contours, hierarchy = cv2.findContours(thresh, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
|
||||
return contours, hierarchy
|
||||
|
||||
def dilate_textline_contours(all_found_textline_polygons):
|
||||
from . import ensure_array
|
||||
return [ensure_array(
|
||||
[polygon2contour(contour2polygon(contour, dilate=6))
|
||||
for contour in region])
|
||||
for region in all_found_textline_polygons]
|
||||
|
||||
def dilate_textregion_contours(all_found_textregion_polygons):
|
||||
from . import ensure_array
|
||||
return ensure_array(
|
||||
[polygon2contour(contour2polygon(contour, dilate=6))
|
||||
for contour in all_found_textregion_polygons])
|
||||
|
||||
def match_deskewed_contours(slope_deskew, contours_o, contours_d, shape_o, shape_d):
|
||||
from . import ensure_array
|
||||
|
||||
cntareas_o = np.array([cv2.contourArea(contour) for contour in contours_o])
|
||||
cntareas_d = np.array([cv2.contourArea(contour) for contour in contours_d])
|
||||
cntareas_o = cntareas_o / float(np.prod(shape_o[:2]))
|
||||
cntareas_d = cntareas_d / float(np.prod(shape_d[:2]))
|
||||
|
||||
contours_o = ensure_array(contours_o)
|
||||
contours_d = ensure_array(contours_d)
|
||||
|
||||
sort_o = np.argsort(cntareas_o)
|
||||
sort_d = np.argsort(cntareas_d)
|
||||
contours_o = contours_o[sort_o]
|
||||
contours_d = contours_d[sort_d]
|
||||
cntareas_o = cntareas_o[sort_o]
|
||||
cntareas_d = cntareas_d[sort_d]
|
||||
|
||||
centers_o = np.stack(find_center_of_contours(contours_o)) # [2, N]
|
||||
centers_d = np.stack(find_center_of_contours(contours_d)) # [2, N]
|
||||
center0_o = centers_o[:, -1:] # [2, 1]
|
||||
center0_d = centers_d[:, -1:] # [2, 1]
|
||||
|
||||
# find the largest among the largest 5 deskewed contours
|
||||
# that is also closest to the largest original contour
|
||||
last5_centers_d = centers_d[:, -5:]
|
||||
dists_d = np.linalg.norm(center0_o - last5_centers_d, axis=0)
|
||||
ind_largest = len(contours_d) - last5_centers_d.shape[1] + np.argmin(dists_d)
|
||||
center0_d[:, 0] = centers_d[:, ind_largest]
|
||||
|
||||
# order new contours the same way as the undeskewed contours
|
||||
# (by calculating the offset of the largest contours, respectively,
|
||||
# of the new and undeskewed image; then for each contour,
|
||||
# finding the closest new contour, with proximity calculated
|
||||
# as distance of their centers modulo offset vector)
|
||||
h_o, w_o = shape_o[:2]
|
||||
center_o = (w_o // 2, h_o // 2)
|
||||
M = cv2.getRotationMatrix2D(center_o, slope_deskew, 1.0)
|
||||
M_22 = np.array(M)[:2, :2]
|
||||
center0_o = np.dot(M_22, center0_o) # [2, 1]
|
||||
offset = center0_o - center0_d # [2, 1]
|
||||
|
||||
centers_o = np.dot(M_22, centers_o) - offset # [2,N]
|
||||
# add dimension for area (so only contours of similar size will be considered close)
|
||||
centers_o = np.append(centers_o, cntareas_o[np.newaxis], axis=0)
|
||||
centers_d = np.append(centers_d, cntareas_d[np.newaxis], axis=0)
|
||||
|
||||
dists = np.zeros((len(contours_o), len(contours_d)))
|
||||
for i in range(len(contours_o)):
|
||||
dists[i] = np.linalg.norm(centers_o[:, i: i + 1] - centers_d, axis=0)
|
||||
corresp = np.zeros(dists.shape, dtype=bool)
|
||||
# keep searching next-closest until at least one correspondence on each side
|
||||
while not np.all(corresp.sum(axis=1)) or not np.all(corresp.sum(axis=0)):
|
||||
idx = np.nanargmin(dists)
|
||||
i, j = np.unravel_index(idx, dists.shape)
|
||||
dists[i, j] = np.nan
|
||||
corresp[i, j] = True
|
||||
# print("original/deskewed adjacency", corresp.nonzero())
|
||||
contours_d_ordered = contours_d[np.argmax(corresp, axis=1)]
|
||||
# from matplotlib import pyplot as plt
|
||||
# img1 = np.zeros(shape_d[:2], dtype=np.uint8)
|
||||
# for i in range(len(contours_o)):
|
||||
# cv2.fillPoly(img1, pts=[contours_d_ordered[i]], color=i + 1)
|
||||
# plt.subplot(1, 4, 1, title="direct corresp contours")
|
||||
# plt.imshow(img1)
|
||||
# img2 = np.zeros(shape_d[:2], dtype=np.uint8)
|
||||
# join deskewed regions mapping to single original ones
|
||||
for i in range(len(contours_o)):
|
||||
if np.count_nonzero(corresp[i]) > 1:
|
||||
indices = np.flatnonzero(corresp[i])
|
||||
# print("joining", indices)
|
||||
polygons_d = [contour2polygon(contour)
|
||||
for contour in contours_d[indices]]
|
||||
contour_d_joined = polygon2contour(join_polygons(polygons_d))
|
||||
contours_d_ordered[i] = contour_d_joined
|
||||
# cv2.fillPoly(img2, pts=[contour_d_joined], color=i + 1)
|
||||
# plt.subplot(1, 4, 2, title="joined contours")
|
||||
# plt.imshow(img2)
|
||||
# img3 = np.zeros(shape_d[:2], dtype=np.uint8)
|
||||
# split deskewed regions mapping to multiple original ones
|
||||
def deskew(polygon):
|
||||
polygon = affinity.rotate(polygon, -slope_deskew, origin=center_o)
|
||||
#polygon = affinity.translate(polygon, *offset.squeeze())
|
||||
return polygon
|
||||
for j in range(len(contours_d)):
|
||||
if np.count_nonzero(corresp[:, j]) > 1:
|
||||
indices = np.flatnonzero(corresp[:, j])
|
||||
# print("splitting along", indices)
|
||||
polygons_o = [deskew(contour2polygon(contour))
|
||||
for contour in contours_o[indices]]
|
||||
polygon_d = contour2polygon(contours_d[j])
|
||||
polygons_d = [make_intersection(polygon_d, polygon)
|
||||
for polygon in polygons_o]
|
||||
# ignore where there is no actual overlap
|
||||
indices = indices[np.flatnonzero(polygons_d)]
|
||||
contours_d_joined = [polygon2contour(polygon_d)
|
||||
for polygon_d in polygons_d
|
||||
if polygon_d]
|
||||
contours_d_ordered[indices] = contours_d_joined
|
||||
# cv2.fillPoly(img3, pts=contours_d_joined, color=j + 1)
|
||||
# plt.subplot(1, 4, 3, title="split contours")
|
||||
# plt.imshow(img3)
|
||||
# img4 = np.zeros(shape_d[:2], dtype=np.uint8)
|
||||
# for i in range(len(contours_o)):
|
||||
# cv2.fillPoly(img4, pts=[contours_d_ordered[i]], color=i + 1)
|
||||
# plt.subplot(1, 4, 4, title="result contours")
|
||||
# plt.imshow(img4)
|
||||
# plt.show()
|
||||
# from matplotlib import patches as ptchs
|
||||
# plt.subplot(1, 2, 1, title="undeskewed")
|
||||
# plt.imshow(mask_o)
|
||||
# centers_o = np.stack(find_center_of_contours(contours_o)) # [2, N]
|
||||
# for i in range(len(contours_o)):
|
||||
# cnt = contours_o[i]
|
||||
# ctr = centers_o[:, i]
|
||||
# plt.gca().add_patch(ptchs.Polygon(cnt[:, 0], closed=False, fill=False, color='blue'))
|
||||
# plt.gca().scatter(ctr[0], ctr[1], 20, c='blue', marker='x')
|
||||
# plt.gca().text(ctr[0], ctr[1], str(i), c='blue')
|
||||
# plt.subplot(1, 2, 2, title="deskewed")
|
||||
# plt.imshow(mask_d)
|
||||
# centers_d = np.stack(find_center_of_contours(contours_d_ordered)) # [2, N]
|
||||
# for i in range(len(contours_o)):
|
||||
# cnt = contours_o[i]
|
||||
# cnt = polygon2contour(deskew(contour2polygon(cnt)))
|
||||
# plt.gca().add_patch(ptchs.Polygon(cnt[:, 0], closed=False, fill=False, color='blue'))
|
||||
# for i in range(len(contours_d_ordered)):
|
||||
# cnt = contours_d_ordered[i]
|
||||
# ctr = centers_d[:, i]
|
||||
# plt.gca().add_patch(ptchs.Polygon(cnt[:, 0], closed=False, fill=False, color='red'))
|
||||
# plt.gca().scatter(ctr[0], ctr[1], 20, c='red', marker='x')
|
||||
# plt.gca().text(ctr[0], ctr[1], str(i), c='red')
|
||||
# plt.show()
|
||||
invsort_o = np.argsort(sort_o)
|
||||
return contours_d_ordered[invsort_o]
|
||||
|
||||
def estimate_skew_contours(contours):
|
||||
if not len(contours):
|
||||
raise ValueError("not enough contours")
|
||||
_, size_in, angle_in = zip(*map(cv2.minAreaRect, contours))
|
||||
w_in, h_in = np.array(size_in).T
|
||||
angle_in = np.array(angle_in)
|
||||
# 1. depending on how contours are oriented,
|
||||
# and where they start, minAreaRect can present
|
||||
# either side as width or height; so we first
|
||||
# need to normalise
|
||||
transposed = h_in > w_in
|
||||
# print("transposed", transposed, angle_in)
|
||||
w_in[transposed], h_in[transposed] = h_in[transposed], w_in[transposed]
|
||||
angle_in[transposed] -= 90
|
||||
# 2. now we look at aspect ratio: too short
|
||||
# textlines do not yield reliable angles
|
||||
usable = w_in > 2.5 * h_in
|
||||
# print("usable aspect", w_in / h_in, usable, angle_in[usable])
|
||||
if not np.any(usable):
|
||||
raise ValueError("not enough contours with high aspect ratio")
|
||||
# 3. next, get rid of outliers regarding length
|
||||
w_avg = np.median(w_in[usable])
|
||||
w_dev = w_in[usable] / w_avg
|
||||
usable[usable] = (0.67 <= w_dev) & (w_dev <= 1.33)
|
||||
# print("usable length", w_in[usable] / w_avg, usable, angle_in[usable])
|
||||
if not np.any(usable):
|
||||
raise ValueError("not enough contours with consistent length")
|
||||
if np.count_nonzero(usable) == 1:
|
||||
return angle_in[usable]
|
||||
# 4. there is no way to distinguish between +90 and -89.9 here,
|
||||
# so map to [0,180] when calculating averages, then map back to [-90,90]
|
||||
# (we don't want -90 and +89 to average zero, or +1 and +179 to average 90)
|
||||
angles = angle_in[usable]
|
||||
if transposed := np.median(np.abs(angles)) >= 45:
|
||||
angles %= 180
|
||||
angle_avg = np.median(angles)
|
||||
angle_dev = np.abs(angles - angle_avg)
|
||||
usable[usable] = (angle_dev <= 2 * np.median(angle_dev))
|
||||
# print("usable angle", usable, angle_in[usable])
|
||||
if not np.any(usable):
|
||||
raise ValueError("not enough contours with consistent angle")
|
||||
if transposed:
|
||||
angle = 90 - (90 - np.mean(angle_in[usable] % 180)) % 180
|
||||
else:
|
||||
angle = np.mean(angle_in[usable])
|
||||
# print("mean angle", angle)
|
||||
return angle
|
||||
|
||||
def contour2polygon(contour: Union[np.ndarray, Sequence[Sequence[Sequence[Number]]]], dilate=0):
|
||||
polygon = Polygon([point[0] for point in contour])
|
||||
if dilate:
|
||||
polygon = polygon.buffer(dilate)
|
||||
if polygon.geom_type == 'GeometryCollection':
|
||||
# heterogeneous result: filter zero-area shapes (LineString, Point)
|
||||
polygon = unary_union([geom for geom in polygon.geoms if geom.area > 0])
|
||||
if polygon.geom_type == 'MultiPolygon':
|
||||
# homogeneous result: construct convex hull to connect
|
||||
polygon = join_polygons(polygon.geoms)
|
||||
return make_valid(polygon)
|
||||
|
||||
def polygon2contour(polygon: Polygon) -> np.ndarray:
|
||||
polygon = np.array(polygon.exterior.coords[:-1], dtype=int)
|
||||
return np.maximum(0, polygon).astype(int)[:, np.newaxis]
|
||||
|
||||
def make_intersection(poly1, poly2):
|
||||
interp = poly1.intersection(poly2)
|
||||
# post-process
|
||||
if interp.is_empty or interp.area == 0.0:
|
||||
return None
|
||||
if interp.geom_type == 'GeometryCollection':
|
||||
# heterogeneous result: filter zero-area shapes (LineString, Point)
|
||||
interp = unary_union([geom for geom in interp.geoms if geom.area > 0])
|
||||
if interp.geom_type == 'MultiPolygon':
|
||||
# homogeneous result: construct convex hull to connect
|
||||
interp = join_polygons(interp.geoms)
|
||||
assert interp.geom_type == 'Polygon', interp.wkt
|
||||
interp = make_valid(interp)
|
||||
return interp
|
||||
|
||||
def make_valid(polygon: Polygon) -> Polygon:
|
||||
"""Ensures shapely.geometry.Polygon object is valid by repeated rearrangement/simplification/enlargement."""
|
||||
def isint(x):
|
||||
return isinstance(x, int) or int(x) == x
|
||||
# make sure rounding does not invalidate
|
||||
if (not all(map(isint, np.array(polygon.exterior.coords).flat)) and
|
||||
polygon.minimum_clearance < 1.0):
|
||||
polygon = Polygon(np.round(polygon.exterior.coords))
|
||||
if polygon.is_valid:
|
||||
return polygon
|
||||
points = list(polygon.exterior.coords[:-1])
|
||||
def step(split, tolerance):
|
||||
# try by re-arranging points
|
||||
poly = Polygon(points[-split:]+points[:-split])
|
||||
if poly.is_valid:
|
||||
return poly
|
||||
# try by simplification
|
||||
poly = poly.simplify(tolerance + 1.0)
|
||||
if poly.is_valid:
|
||||
return poly
|
||||
# try by enlarging
|
||||
poly = poly.buffer(tolerance)
|
||||
if poly.is_valid:
|
||||
return poly
|
||||
return None
|
||||
for split in range(1, len(points)):
|
||||
for tolerance in np.linspace(1, np.sqrt(polygon.area), 100):
|
||||
# simplification may not be possible (at all) due to ordering
|
||||
# in that case, try another starting point
|
||||
if poly := step(split, tolerance):
|
||||
return poly
|
||||
assert polygon.is_valid, polygon.wkt
|
||||
return polygon
|
||||
|
||||
def join_polygons(polygons: Sequence[Polygon], scale=20) -> Polygon:
|
||||
"""construct concave hull (alpha shape) from input polygons by connecting their pairwise nearest points"""
|
||||
# ensure input polygons are simply typed and all oriented equally
|
||||
polygons = [orient(poly)
|
||||
for poly in itertools.chain.from_iterable(
|
||||
[poly.geoms
|
||||
if poly.geom_type in ['MultiPolygon', 'GeometryCollection']
|
||||
else [poly]
|
||||
for poly in polygons])]
|
||||
npoly = len(polygons)
|
||||
if npoly == 1:
|
||||
return polygons[0]
|
||||
# find min-dist path through all polygons (travelling salesman)
|
||||
pairs = itertools.combinations(range(npoly), 2)
|
||||
dists = np.zeros((npoly, npoly), dtype=float)
|
||||
for i, j in pairs:
|
||||
dist = polygons[i].distance(polygons[j])
|
||||
if dist < 1e-5:
|
||||
dist = 1e-5 # if pair merely touches, we still need to get an edge
|
||||
dists[i, j] = dist
|
||||
dists[j, i] = dist
|
||||
dists = minimum_spanning_tree(dists, overwrite=True)
|
||||
# add bridge polygons (where necessary)
|
||||
for prevp, nextp in zip(*dists.nonzero()):
|
||||
prevp = polygons[prevp]
|
||||
nextp = polygons[nextp]
|
||||
nearest = nearest_points(prevp, nextp)
|
||||
bridgep = orient(LineString(nearest).buffer(max(1, scale/5), resolution=1), -1)
|
||||
polygons.append(bridgep)
|
||||
jointp = unary_union(polygons)
|
||||
if jointp.geom_type == 'MultiPolygon':
|
||||
jointp = unary_union(jointp.geoms)
|
||||
assert jointp.geom_type == 'Polygon', jointp.wkt
|
||||
# follow-up calculations will necessarily be integer;
|
||||
# so anticipate rounding here and then ensure validity
|
||||
jointp2 = set_precision(jointp, 1.0, mode="keep_collapsed")
|
||||
if jointp2.geom_type != 'Polygon' or not jointp2.is_valid:
|
||||
jointp2 = Polygon(np.round(jointp.exterior.coords))
|
||||
jointp2 = make_valid(jointp2)
|
||||
assert jointp2.geom_type == 'Polygon', jointp2.wkt
|
||||
return jointp2
|
||||
|
|
@ -1,16 +0,0 @@
|
|||
|
||||
# cannot use importlib.resources until we move to 3.9+ forimportlib.resources.files
|
||||
import sys
|
||||
from PIL import ImageFont
|
||||
|
||||
if sys.version_info < (3, 10):
|
||||
import importlib_resources
|
||||
else:
|
||||
import importlib.resources as importlib_resources
|
||||
|
||||
|
||||
def get_font():
|
||||
#font_path = "Charis-7.000/Charis-Regular.ttf" # Make sure this file exists!
|
||||
font = importlib_resources.files(__package__) / "../Charis-Regular.ttf"
|
||||
with importlib_resources.as_file(font) as font:
|
||||
return ImageFont.truetype(font=font, size=40)
|
||||
|
|
@ -1,185 +0,0 @@
|
|||
import numpy as np
|
||||
import cv2
|
||||
from scipy.signal import find_peaks
|
||||
from scipy.ndimage import gaussian_filter1d
|
||||
from .contour import find_center_of_contours, return_contours_of_interested_region
|
||||
from .resize import resize_image
|
||||
from .rotate import rotate_image
|
||||
|
||||
def get_marginals(num_col, slope_deskew, early_layout,
|
||||
kernel=None,
|
||||
label_text=1,
|
||||
label_marg=4,
|
||||
label_tabs=10,
|
||||
):
|
||||
if kernel is None:
|
||||
kernel = np.ones((5, 5), dtype=np.uint8)
|
||||
kernel_hor = np.ones((1, 5), dtype=np.uint8)
|
||||
|
||||
text_mask = ((early_layout == label_text) |
|
||||
(early_layout == label_tabs)).astype(np.uint8)
|
||||
text_mask_d = rotate_image(text_mask, slope_deskew)
|
||||
main_mask_d = np.zeros_like(text_mask_d)
|
||||
height, width = main_mask_d.shape
|
||||
|
||||
if height <= 1500:
|
||||
pass
|
||||
elif 1500 < height <= 1800:
|
||||
text_mask_d = resize_image(text_mask_d, int(height / 1.5), width)
|
||||
text_mask_d = cv2.erode(text_mask_d, kernel, iterations=5)
|
||||
# rs: and back to original size
|
||||
text_mask_d = resize_image(text_mask_d, height, width)
|
||||
else:
|
||||
text_mask_d = resize_image(text_mask_d, int(height / 1.8), width)
|
||||
text_mask_d = cv2.erode(text_mask_d, kernel, iterations=7)
|
||||
# rs: and back to original size
|
||||
text_mask_d = resize_image(text_mask_d, height, width)
|
||||
|
||||
text_mask_d = cv2.erode(text_mask_d, kernel_hor, iterations=6)
|
||||
text_mask_d_y = text_mask_d.sum(axis=0)
|
||||
text_mask_d_y_eroded = text_mask_d.sum(axis=0)
|
||||
|
||||
max_text_thickness_percent = 100. * text_mask_d_y.max() / height
|
||||
min_text_thickness = max_text_thickness_percent / 100. * height / 20.
|
||||
|
||||
# plt.figure()
|
||||
# ax1 = plt.subplot(2, 2, 1, title="text_mask_d")
|
||||
# ax1.imshow(text_mask_d, aspect='auto')
|
||||
# ax2 = plt.subplot(2, 2, 3, title="text_mask_d_y", sharex=ax1)
|
||||
# ax2.plot(list(range(width)), text_mask_d_y)
|
||||
# ax2.hlines(int(0.14 * height), 0, width,
|
||||
# label='max_text_thickness=14%', colors='r')
|
||||
# ax2.hlines([min_text_thickness], 0, width,
|
||||
# label='min_text_thickness', colors='g')
|
||||
# ax2.scatter([np.argmax(text_mask_d_y)],
|
||||
# [text_mask_d_y.max()], color='r',
|
||||
# label='max = %d%%' % max_text_thickness_percent)
|
||||
# ax1 = plt.subplot(2, 2, 4, title="early layout")
|
||||
# ax1.imshow(early_layout, aspect='auto')
|
||||
# plt.legend()
|
||||
# plt.show()
|
||||
|
||||
if max_text_thickness_percent < 14:
|
||||
return
|
||||
|
||||
text_mask_d_y_rev = np.max(text_mask_d_y) - text_mask_d_y
|
||||
region_sum_0 = gaussian_filter1d(text_mask_d_y, 1)
|
||||
first_nonzero = region_sum_0.nonzero()[0][0] # outer left
|
||||
last_nonzero = region_sum_0.nonzero()[0][-1] # outer right
|
||||
mid_point = 0.5 * (last_nonzero + first_nonzero)
|
||||
one_third_right = (last_nonzero - mid_point) / 3.0
|
||||
one_third_left = (mid_point - first_nonzero) / 3.0
|
||||
|
||||
# rs: constrain the distance at least 2 characters at 12pt, retrieve height and prominence
|
||||
peaks, props = find_peaks(text_mask_d_y_rev, height=0, prominence=0, distance=30)
|
||||
peaks_orig = np.copy(peaks)
|
||||
# rs: also calculate the product of prominence and height (for final selection)
|
||||
scores = np.zeros(peaks.max() + 1)
|
||||
scores[peaks] = props['prominences'] * props['peak_heights']
|
||||
|
||||
peaks = peaks[(peaks > first_nonzero) & (peaks < last_nonzero)]
|
||||
peaks = peaks[region_sum_0[peaks] < min_text_thickness]
|
||||
|
||||
if num_col == 1:
|
||||
peaks_right = peaks[peaks > mid_point]
|
||||
peaks_left = peaks[peaks < mid_point]
|
||||
elif num_col == 2:
|
||||
peaks_right = peaks[peaks > mid_point + one_third_right]
|
||||
peaks_left = peaks[peaks < mid_point - one_third_left]
|
||||
else:
|
||||
# should not happen, anyway
|
||||
return
|
||||
|
||||
if len(peaks_left) == 0:
|
||||
if len(peaks_right) == 0:
|
||||
# plt.figure()
|
||||
# ax1 = plt.subplot(2, 1, 1, title='text_mask_d (deskewed text+sep mask)')
|
||||
# ax1.imshow(text_mask_d, aspect='auto')
|
||||
# ax1.vlines([first_nonzero], 0, height, label='first_nonzero', colors='r')
|
||||
# ax1.vlines([last_nonzero], 0, height, label='last_nonzero', colors='r')
|
||||
# ax1.vlines(peaks_left, 0, height, label='peaks_left', colors='orange')
|
||||
# ax1.vlines(peaks_right, 0, height, label='peaks_right', colors='orange')
|
||||
# ax2 = plt.subplot(2, 1, 2, title='text_mask_d_y (smoothed)', sharex=ax1)
|
||||
# ax2.plot(list(range(width)), region_sum_0)
|
||||
# ax2.hlines(min_text_thickness, 0, width, colors='g',
|
||||
# label='min_text_thickness=%d' % min_text_thickness)
|
||||
# ax2.scatter(peaks_orig, region_sum_0[peaks_orig], label='peaks')
|
||||
# plt.legend()
|
||||
# plt.show()
|
||||
return
|
||||
point_right = peaks_right[np.argmax(scores[peaks_right])]
|
||||
#point_left = first_nonzero
|
||||
point_left = 0
|
||||
elif len(peaks_right) == 0:
|
||||
point_left = peaks_left[np.argmax(scores[peaks_left])]
|
||||
#point_right = last_nonzero
|
||||
point_right = width - 1
|
||||
else:
|
||||
best_left = np.argmax(scores[peaks_left])
|
||||
best_right = np.argmax(scores[peaks_right])
|
||||
point_left = peaks_left[best_left]
|
||||
point_right = peaks_right[best_right]
|
||||
if scores[best_left] < 0.1 * scores[best_right]:
|
||||
point_left = 0
|
||||
#point_left = first_nonzero
|
||||
if scores[best_right] < 0.1 * scores[best_left]:
|
||||
point_right = 0
|
||||
#point_right = last_nonzero
|
||||
|
||||
main_mask_d[:, point_left: point_right] = 1
|
||||
if not np.any(main_mask_d):
|
||||
return
|
||||
|
||||
# plt.figure()
|
||||
# ax1 = plt.subplot(2, 2, 1)
|
||||
# ax1.title.set_text('text_mask_d (deskewed text+table mask)')
|
||||
# ax1.imshow(text_mask_d)
|
||||
# ax1.vlines(peaks_left, 0, height, label='peaks_left', colors='b')
|
||||
# ax1.vlines(peaks_right, 0, height, label='peaks_right', colors='b')
|
||||
# ax1.vlines([first_nonzero], 0, height, label='first_nonzero', colors='g')
|
||||
# ax1.vlines([last_nonzero], 0, height, label='last_nonzero', colors='g')
|
||||
# ax1.vlines([point_left], 0, height, label='point_left', colors='r')
|
||||
# ax1.vlines([point_right], 0, height, label='point_right', colors='r')
|
||||
# ax2 = plt.subplot(2, 2, 2, title='main_mask_d (deskewed main mask)', sharey=ax1)
|
||||
# ax2.imshow(main_mask_d)
|
||||
# ax3 = plt.subplot(2, 2, 3, title='text_mask_d_y (projection for minima)', sharex=ax1)
|
||||
# ax3.plot(list(range(width)), text_mask_d_y)
|
||||
# ax3.set_aspect('auto')
|
||||
# ax4 = plt.subplot(2, 2, 4, title='early_layout (undeskewed labels)')
|
||||
# ax4.imshow(early_layout)
|
||||
# plt.legend()
|
||||
# plt.show()
|
||||
|
||||
# rs: rotate back (into undeskewed/original shape as early_layout input):
|
||||
main_mask = rotate_image(main_mask_d, -slope_deskew)
|
||||
|
||||
min_area_text = 0.00001
|
||||
main_contour = return_contours_of_interested_region(main_mask, 1, min_area_text)[0]
|
||||
text_contours = return_contours_of_interested_region(early_layout, label_text, min_area_text)
|
||||
cx_text, cy_text = find_center_of_contours(text_contours)
|
||||
|
||||
marg_contours = []
|
||||
for i, contour in enumerate(text_contours):
|
||||
if -1 == cv2.pointPolygonTest(main_contour,
|
||||
(cx_text[i],
|
||||
cy_text[i]),
|
||||
False):
|
||||
marg_contours.append(contour)
|
||||
|
||||
# early_layout_orig = np.copy(early_layout)
|
||||
early_layout = cv2.fillPoly(early_layout, pts=marg_contours, color=label_marg)
|
||||
|
||||
# plt.figure()
|
||||
# ax1 = plt.subplot(2, 2, 1, title='main_mask_d (deskewed main mask)')
|
||||
# plt.imshow(main_mask_d)
|
||||
# ax2 = plt.subplot(2, 2, 2, title='main_mask (undeskewed main mask)')
|
||||
# plt.imshow(main_mask)
|
||||
# ax3 = plt.subplot(2, 2, 3, title='early_layout (undeskewed labels original)')
|
||||
# plt.imshow(early_layout_orig)
|
||||
# ax4 = plt.subplot(2, 2, 4, title='early_layout (undeskewed labels split)')
|
||||
# plt.imshow(early_layout)
|
||||
# plt.show()
|
||||
|
||||
# if there was no main text, then relabel marginalia as main
|
||||
if not np.any(early_layout == label_text):
|
||||
early_layout[early_layout == label_marg] = label_text
|
||||
|
|
@ -1,38 +0,0 @@
|
|||
import math
|
||||
import cv2
|
||||
|
||||
|
||||
def rotation_image_new(img, thetha):
|
||||
rotated = rotate_image(img, thetha)
|
||||
return rotate_max_area_new(img, rotated, thetha)
|
||||
|
||||
def rotate_image(img_patch, slope):
|
||||
(h, w) = img_patch.shape[:2]
|
||||
center = (w // 2, h // 2)
|
||||
M = cv2.getRotationMatrix2D(center, slope, 1.0)
|
||||
return cv2.warpAffine(img_patch, M, (w, h) )
|
||||
|
||||
def rotate_image_different( img, slope):
|
||||
# img = cv2.imread('images/input.jpg')
|
||||
num_rows, num_cols = img.shape[:2]
|
||||
|
||||
rotation_matrix = cv2.getRotationMatrix2D((num_cols / 2, num_rows / 2), slope, 1)
|
||||
img_rotation = cv2.warpAffine(img, rotation_matrix, (num_cols, num_rows))
|
||||
return img_rotation
|
||||
|
||||
def rotate_image_enlarge(img, angle):
|
||||
h, w = img.shape[:2]
|
||||
cx, cy = 0.5 * w, 0.5 * h
|
||||
matrix = cv2.getRotationMatrix2D((cx, cy), angle, 1.0)
|
||||
radian = angle / 180 * math.pi
|
||||
cos = abs(math.cos(radian))
|
||||
sin = abs(math.sin(radian))
|
||||
new_w, new_h = (w * cos + h * sin,
|
||||
w * sin + h * cos)
|
||||
# box is larger after resize, so instead of shifting
|
||||
# back from center, shift from new center
|
||||
matrix[0, 2] += 0.5 * new_w - cx
|
||||
matrix[1, 2] += 0.5 * new_h - cy
|
||||
return cv2.warpAffine(img, matrix, (int(new_w + 0.5),
|
||||
int(new_h + 0.5)),
|
||||
flags=cv2.INTER_CUBIC)
|
||||
|
|
@ -1,45 +0,0 @@
|
|||
from multiprocessing import shared_memory
|
||||
from contextlib import contextmanager
|
||||
from functools import wraps
|
||||
import numpy as np
|
||||
|
||||
@contextmanager
|
||||
def share_ndarray(array: np.ndarray):
|
||||
size = np.dtype(array.dtype).itemsize * np.prod(array.shape)
|
||||
shm = shared_memory.SharedMemory(create=True, size=size)
|
||||
try:
|
||||
shared_array = np.ndarray(array.shape, dtype=array.dtype, buffer=shm.buf)
|
||||
shared_array[:] = array[:]
|
||||
shared_array.flags["WRITEABLE"] = False
|
||||
yield dict(shape=array.shape, dtype=array.dtype, name=shm.name)
|
||||
finally:
|
||||
shm.close()
|
||||
shm.unlink()
|
||||
|
||||
@contextmanager
|
||||
def ndarray_shared(array: dict):
|
||||
shm = shared_memory.SharedMemory(name=array['name'])
|
||||
try:
|
||||
array = np.ndarray(array['shape'], dtype=array['dtype'], buffer=shm.buf)
|
||||
yield array
|
||||
finally:
|
||||
shm.close()
|
||||
|
||||
def wrap_ndarray_shared(kw=None):
|
||||
def wrapper(f):
|
||||
if kw is None:
|
||||
@wraps(f)
|
||||
def shared_func(array, *args, **kwargs):
|
||||
with ndarray_shared(array) as ndarray:
|
||||
return f(ndarray, *args, **kwargs)
|
||||
return shared_func
|
||||
else:
|
||||
@wraps(f)
|
||||
def shared_func(*args, **kwargs):
|
||||
array = kwargs.pop(kw)
|
||||
with ndarray_shared(array) as ndarray:
|
||||
kwargs[kw] = ndarray
|
||||
return f(*args, **kwargs)
|
||||
return shared_func
|
||||
return wrapper
|
||||
|
||||
|
|
@ -1,504 +0,0 @@
|
|||
import math
|
||||
import copy
|
||||
|
||||
import numpy as np
|
||||
import cv2
|
||||
import tensorflow as tf
|
||||
from scipy.signal import find_peaks
|
||||
from scipy.ndimage import gaussian_filter1d
|
||||
from PIL import Image, ImageDraw, ImageFont
|
||||
|
||||
from .resize import resize_image
|
||||
|
||||
|
||||
def decode_batch_predictions(pred, num_to_char, max_len = 128):
|
||||
# input_len is the product of the batch size and the
|
||||
# number of time steps.
|
||||
input_len = np.ones(pred.shape[0]) * pred.shape[1]
|
||||
|
||||
# Decode CTC predictions using greedy search.
|
||||
# decoded is a tuple with 2 elements.
|
||||
decoded = tf.keras.backend.ctc_decode(pred,
|
||||
input_length = input_len,
|
||||
beam_width = 100)
|
||||
# The outputs are in the first element of the tuple.
|
||||
# Additionally, the first element is actually a list,
|
||||
# therefore we take the first element of that list as well.
|
||||
#print(decoded,'decoded')
|
||||
decoded = decoded[0][0][:, :max_len]
|
||||
|
||||
#print(decoded, decoded.shape,'decoded')
|
||||
|
||||
output = []
|
||||
for d in decoded:
|
||||
# Convert the predicted indices to the corresponding chars.
|
||||
d = tf.strings.reduce_join(num_to_char(d))
|
||||
d = d.numpy().decode("utf-8")
|
||||
output.append(d)
|
||||
return output
|
||||
|
||||
|
||||
def distortion_free_resize(image, img_size):
|
||||
w, h = img_size
|
||||
image = tf.image.resize(image, size=(h, w), preserve_aspect_ratio=True)
|
||||
|
||||
# Check tha amount of padding needed to be done.
|
||||
pad_height = h - tf.shape(image)[0]
|
||||
pad_width = w - tf.shape(image)[1]
|
||||
|
||||
# Only necessary if you want to do same amount of padding on both sides.
|
||||
if pad_height % 2 != 0:
|
||||
height = pad_height // 2
|
||||
pad_height_top = height + 1
|
||||
pad_height_bottom = height
|
||||
else:
|
||||
pad_height_top = pad_height_bottom = pad_height // 2
|
||||
|
||||
if pad_width % 2 != 0:
|
||||
width = pad_width // 2
|
||||
pad_width_left = width + 1
|
||||
pad_width_right = width
|
||||
else:
|
||||
pad_width_left = pad_width_right = pad_width // 2
|
||||
|
||||
image = tf.pad(
|
||||
image,
|
||||
paddings=[
|
||||
[pad_height_top, pad_height_bottom],
|
||||
[pad_width_left, pad_width_right],
|
||||
[0, 0],
|
||||
],
|
||||
)
|
||||
|
||||
image = tf.transpose(image, (1, 0, 2))
|
||||
image = tf.image.flip_left_right(image)
|
||||
return image
|
||||
|
||||
def return_start_and_end_of_common_text_of_textline_ocr_without_common_section(textline_image):
|
||||
width = np.shape(textline_image)[1]
|
||||
height = np.shape(textline_image)[0]
|
||||
common_window = int(0.06*width)
|
||||
|
||||
width1 = int ( width/2. - common_window )
|
||||
width2 = int ( width/2. + common_window )
|
||||
|
||||
img_sum = np.sum(textline_image[:,:,0], axis=0)
|
||||
sum_smoothed = gaussian_filter1d(img_sum, 3)
|
||||
|
||||
peaks_real, _ = find_peaks(sum_smoothed, height=0)
|
||||
if len(peaks_real)>70:
|
||||
|
||||
peaks_real = peaks_real[(peaks_real<width2) & (peaks_real>width1)]
|
||||
|
||||
arg_max = np.argmax(sum_smoothed[peaks_real])
|
||||
peaks_final = peaks_real[arg_max]
|
||||
return peaks_final
|
||||
else:
|
||||
return None
|
||||
|
||||
# Function to fit text inside the given area
|
||||
def fit_text_single_line(draw, text, font_path, max_width, max_height):
|
||||
initial_font_size = 50
|
||||
font_size = initial_font_size
|
||||
while font_size > 10: # Minimum font size
|
||||
font = ImageFont.truetype(font_path, font_size)
|
||||
text_bbox = draw.textbbox((0, 0), text, font=font) # Get text bounding box
|
||||
text_width = text_bbox[2] - text_bbox[0]
|
||||
text_height = text_bbox[3] - text_bbox[1]
|
||||
|
||||
if text_width <= max_width and text_height <= max_height:
|
||||
return font # Return the best-fitting font
|
||||
|
||||
font_size -= 2 # Reduce font size and retry
|
||||
|
||||
return ImageFont.truetype(font_path, 10) # Smallest font fallback
|
||||
|
||||
def return_textlines_split_if_needed(textline_image, textline_image_bin=None):
|
||||
|
||||
split_point = return_start_and_end_of_common_text_of_textline_ocr_without_common_section(textline_image)
|
||||
if split_point:
|
||||
image1 = textline_image[:, :split_point,:]# image.crop((0, 0, width2, height))
|
||||
image2 = textline_image[:, split_point:,:]#image.crop((width1, 0, width, height))
|
||||
if textline_image_bin is not None:
|
||||
image1_bin = textline_image_bin[:, :split_point,:]# image.crop((0, 0, width2, height))
|
||||
image2_bin = textline_image_bin[:, split_point:,:]#image.crop((width1, 0, width, height))
|
||||
return [image1, image2], [image1_bin, image2_bin]
|
||||
else:
|
||||
return [image1, image2], None
|
||||
else:
|
||||
return None, None
|
||||
|
||||
def preprocess_and_resize_image_for_ocrcnn_model(img, image_height, image_width):
|
||||
if img.shape[0]==0 or img.shape[1]==0:
|
||||
img_fin = np.ones((image_height, image_width, 3))
|
||||
else:
|
||||
ratio = image_height /float(img.shape[0])
|
||||
w_ratio = int(ratio * img.shape[1])
|
||||
|
||||
if w_ratio <= image_width:
|
||||
width_new = w_ratio
|
||||
else:
|
||||
width_new = image_width
|
||||
|
||||
if width_new == 0:
|
||||
width_new = img.shape[1]
|
||||
|
||||
|
||||
img = resize_image(img, image_height, width_new)
|
||||
img_fin = np.ones((image_height, image_width, 3))*255
|
||||
|
||||
img_fin[:,:width_new,:] = img[:,:,:]
|
||||
img_fin = img_fin / 255.
|
||||
return img_fin
|
||||
|
||||
def get_deskewed_contour_and_bb_and_image(contour, image, deskew_angle):
|
||||
(h_in, w_in) = image.shape[:2]
|
||||
center = (w_in // 2, h_in // 2)
|
||||
|
||||
rotation_matrix = cv2.getRotationMatrix2D(center, deskew_angle, 1.0)
|
||||
|
||||
cos_angle = abs(rotation_matrix[0, 0])
|
||||
sin_angle = abs(rotation_matrix[0, 1])
|
||||
new_w = int((h_in * sin_angle) + (w_in * cos_angle))
|
||||
new_h = int((h_in * cos_angle) + (w_in * sin_angle))
|
||||
|
||||
rotation_matrix[0, 2] += (new_w / 2) - center[0]
|
||||
rotation_matrix[1, 2] += (new_h / 2) - center[1]
|
||||
|
||||
deskewed_image = cv2.warpAffine(image, rotation_matrix, (new_w, new_h))
|
||||
|
||||
contour_points = np.array(contour, dtype=np.float32)
|
||||
transformed_points = cv2.transform(np.array([contour_points]), rotation_matrix)[0]
|
||||
|
||||
x, y, w, h = cv2.boundingRect(np.array(transformed_points, dtype=np.int32))
|
||||
cropped_textline = deskewed_image[y:y+h, x:x+w]
|
||||
|
||||
return cropped_textline
|
||||
|
||||
def rotate_image_with_padding(image, angle, border_value=(0,0,0)):
|
||||
# Get image dimensions
|
||||
(h, w) = image.shape[:2]
|
||||
|
||||
# Calculate the center of the image
|
||||
center = (w // 2, h // 2)
|
||||
|
||||
# Get the rotation matrix
|
||||
rotation_matrix = cv2.getRotationMatrix2D(center, angle, 1.0)
|
||||
|
||||
# Compute the new bounding dimensions
|
||||
cos = abs(rotation_matrix[0, 0])
|
||||
sin = abs(rotation_matrix[0, 1])
|
||||
new_w = int((h * sin) + (w * cos))
|
||||
new_h = int((h * cos) + (w * sin))
|
||||
|
||||
# Adjust the rotation matrix to account for translation
|
||||
rotation_matrix[0, 2] += (new_w / 2) - center[0]
|
||||
rotation_matrix[1, 2] += (new_h / 2) - center[1]
|
||||
|
||||
# Perform the rotation
|
||||
try:
|
||||
rotated_image = cv2.warpAffine(image, rotation_matrix, (new_w, new_h), borderValue=border_value)
|
||||
except:
|
||||
rotated_image = np.copy(image)
|
||||
|
||||
return rotated_image
|
||||
|
||||
def get_orientation_moments(contour):
|
||||
moments = cv2.moments(contour)
|
||||
if moments["mu20"] - moments["mu02"] == 0: # Avoid division by zero
|
||||
return 90 if moments["mu11"] > 0 else -90
|
||||
else:
|
||||
angle = 0.5 * np.arctan2(2 * moments["mu11"], moments["mu20"] - moments["mu02"])
|
||||
return np.degrees(angle) # Convert radians to degrees
|
||||
|
||||
|
||||
def get_orientation_moments_of_mask(mask):
|
||||
mask=mask.astype('uint8')
|
||||
contours, _ = cv2.findContours(mask[:,:,0], cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
|
||||
|
||||
largest_contour = max(contours, key=cv2.contourArea) if contours else None
|
||||
|
||||
moments = cv2.moments(largest_contour)
|
||||
if moments["mu20"] - moments["mu02"] == 0: # Avoid division by zero
|
||||
return 90 if moments["mu11"] > 0 else -90
|
||||
else:
|
||||
angle = 0.5 * np.arctan2(2 * moments["mu11"], moments["mu20"] - moments["mu02"])
|
||||
return np.degrees(angle) # Convert radians to degrees
|
||||
|
||||
def get_contours_and_bounding_boxes(mask):
|
||||
# Find contours in the binary mask
|
||||
contours, _ = cv2.findContours(mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
|
||||
|
||||
largest_contour = max(contours, key=cv2.contourArea) if contours else None
|
||||
|
||||
# Get the bounding rectangle for the contour
|
||||
x, y, w, h = cv2.boundingRect(largest_contour)
|
||||
#bounding_boxes.append((x, y, w, h))
|
||||
|
||||
return x, y, w, h
|
||||
|
||||
def return_splitting_point_of_image(image_to_spliited):
|
||||
width = np.shape(image_to_spliited)[1]
|
||||
height = np.shape(image_to_spliited)[0]
|
||||
common_window = int(0.03*width)
|
||||
|
||||
width1 = int ( common_window)
|
||||
width2 = int ( width - common_window )
|
||||
|
||||
img_sum = np.sum(image_to_spliited[:,:,0], axis=0)
|
||||
sum_smoothed = gaussian_filter1d(img_sum, 1)
|
||||
|
||||
peaks_real, _ = find_peaks(sum_smoothed, height=0)
|
||||
peaks_real = peaks_real[(peaks_real<width2) & (peaks_real>width1)]
|
||||
|
||||
arg_sort = np.argsort(sum_smoothed[peaks_real])
|
||||
peaks_sort_4 = peaks_real[arg_sort][::-1][:3]
|
||||
|
||||
return np.sort(peaks_sort_4)
|
||||
|
||||
def break_curved_line_into_small_pieces_and_then_merge(img_curved, mask_curved, img_bin_curved=None):
|
||||
peaks_4 = return_splitting_point_of_image(img_curved)
|
||||
if len(peaks_4)>0:
|
||||
imgs_tot = []
|
||||
|
||||
for ind in range(len(peaks_4)+1):
|
||||
if ind==0:
|
||||
img = img_curved[:, :peaks_4[ind], :]
|
||||
if img_bin_curved is not None:
|
||||
img_bin = img_bin_curved[:, :peaks_4[ind], :]
|
||||
mask = mask_curved[:, :peaks_4[ind], :]
|
||||
elif ind==len(peaks_4):
|
||||
img = img_curved[:, peaks_4[ind-1]:, :]
|
||||
if img_bin_curved is not None:
|
||||
img_bin = img_bin_curved[:, peaks_4[ind-1]:, :]
|
||||
mask = mask_curved[:, peaks_4[ind-1]:, :]
|
||||
else:
|
||||
img = img_curved[:, peaks_4[ind-1]:peaks_4[ind], :]
|
||||
if img_bin_curved is not None:
|
||||
img_bin = img_bin_curved[:, peaks_4[ind-1]:peaks_4[ind], :]
|
||||
mask = mask_curved[:, peaks_4[ind-1]:peaks_4[ind], :]
|
||||
|
||||
or_ma = get_orientation_moments_of_mask(mask)
|
||||
|
||||
if img_bin_curved is not None:
|
||||
imgs_tot.append([img, mask, or_ma, img_bin] )
|
||||
else:
|
||||
imgs_tot.append([img, mask, or_ma] )
|
||||
|
||||
|
||||
w_tot_des_list = []
|
||||
w_tot_des = 0
|
||||
imgs_deskewed_list = []
|
||||
imgs_bin_deskewed_list = []
|
||||
|
||||
for ind in range(len(imgs_tot)):
|
||||
img_in = imgs_tot[ind][0]
|
||||
mask_in = imgs_tot[ind][1]
|
||||
ori_in = imgs_tot[ind][2]
|
||||
if img_bin_curved is not None:
|
||||
img_bin_in = imgs_tot[ind][3]
|
||||
|
||||
if abs(ori_in)<45:
|
||||
img_in_des = rotate_image_with_padding(img_in, ori_in, border_value=(255,255,255) )
|
||||
if img_bin_curved is not None:
|
||||
img_bin_in_des = rotate_image_with_padding(img_bin_in, ori_in, border_value=(255,255,255) )
|
||||
mask_in_des = rotate_image_with_padding(mask_in, ori_in)
|
||||
mask_in_des = mask_in_des.astype('uint8')
|
||||
|
||||
#new bounding box
|
||||
x_n, y_n, w_n, h_n = get_contours_and_bounding_boxes(mask_in_des[:,:,0])
|
||||
|
||||
if w_n==0 or h_n==0:
|
||||
img_in_des = np.copy(img_in)
|
||||
if img_bin_curved is not None:
|
||||
img_bin_in_des = np.copy(img_bin_in)
|
||||
w_relative = int(32 * img_in_des.shape[1]/float(img_in_des.shape[0]) )
|
||||
if w_relative==0:
|
||||
w_relative = img_in_des.shape[1]
|
||||
img_in_des = resize_image(img_in_des, 32, w_relative)
|
||||
if img_bin_curved is not None:
|
||||
img_bin_in_des = resize_image(img_bin_in_des, 32, w_relative)
|
||||
else:
|
||||
mask_in_des = mask_in_des[y_n:y_n+h_n, x_n:x_n+w_n, :]
|
||||
img_in_des = img_in_des[y_n:y_n+h_n, x_n:x_n+w_n, :]
|
||||
if img_bin_curved is not None:
|
||||
img_bin_in_des = img_bin_in_des[y_n:y_n+h_n, x_n:x_n+w_n, :]
|
||||
|
||||
w_relative = int(32 * img_in_des.shape[1]/float(img_in_des.shape[0]) )
|
||||
if w_relative==0:
|
||||
w_relative = img_in_des.shape[1]
|
||||
img_in_des = resize_image(img_in_des, 32, w_relative)
|
||||
if img_bin_curved is not None:
|
||||
img_bin_in_des = resize_image(img_bin_in_des, 32, w_relative)
|
||||
|
||||
|
||||
else:
|
||||
img_in_des = np.copy(img_in)
|
||||
if img_bin_curved is not None:
|
||||
img_bin_in_des = np.copy(img_bin_in)
|
||||
w_relative = int(32 * img_in_des.shape[1]/float(img_in_des.shape[0]) )
|
||||
if w_relative==0:
|
||||
w_relative = img_in_des.shape[1]
|
||||
img_in_des = resize_image(img_in_des, 32, w_relative)
|
||||
if img_bin_curved is not None:
|
||||
img_bin_in_des = resize_image(img_bin_in_des, 32, w_relative)
|
||||
|
||||
w_tot_des+=img_in_des.shape[1]
|
||||
w_tot_des_list.append(img_in_des.shape[1])
|
||||
imgs_deskewed_list.append(img_in_des)
|
||||
if img_bin_curved is not None:
|
||||
imgs_bin_deskewed_list.append(img_bin_in_des)
|
||||
|
||||
|
||||
|
||||
|
||||
img_final_deskewed = np.zeros((32, w_tot_des, 3))+255
|
||||
if img_bin_curved is not None:
|
||||
img_bin_final_deskewed = np.zeros((32, w_tot_des, 3))+255
|
||||
else:
|
||||
img_bin_final_deskewed = None
|
||||
|
||||
w_indexer = 0
|
||||
for ind in range(len(w_tot_des_list)):
|
||||
img_final_deskewed[:,w_indexer:w_indexer+w_tot_des_list[ind],:] = imgs_deskewed_list[ind][:,:,:]
|
||||
if img_bin_curved is not None:
|
||||
img_bin_final_deskewed[:,w_indexer:w_indexer+w_tot_des_list[ind],:] = imgs_bin_deskewed_list[ind][:,:,:]
|
||||
w_indexer = w_indexer+w_tot_des_list[ind]
|
||||
return img_final_deskewed, img_bin_final_deskewed
|
||||
else:
|
||||
return img_curved, img_bin_curved
|
||||
|
||||
def return_textline_contour_with_added_box_coordinate(textline_contour, box_ind):
|
||||
textline_contour[:,:,0] += box_ind[2]
|
||||
textline_contour[:,:,1] += box_ind[0]
|
||||
return textline_contour
|
||||
|
||||
|
||||
def return_rnn_cnn_ocr_of_given_textlines(image,
|
||||
all_found_textline_polygons,
|
||||
all_box_coord,
|
||||
prediction_model,
|
||||
b_s_ocr, num_to_char,
|
||||
curved_line=False):
|
||||
max_len = 512
|
||||
padding_token = 299
|
||||
image_width = 512#max_len * 4
|
||||
image_height = 32
|
||||
ind_tot = 0
|
||||
#cv2.imwrite('./img_out.png', image_page)
|
||||
ocr_all_textlines = []
|
||||
cropped_lines_region_indexer = []
|
||||
cropped_lines_meging_indexing = []
|
||||
cropped_lines = []
|
||||
indexer_text_region = 0
|
||||
|
||||
for indexing, ind_poly_first in enumerate(all_found_textline_polygons):
|
||||
#ocr_textline_in_textregion = []
|
||||
if len(ind_poly_first)==0:
|
||||
cropped_lines_region_indexer.append(indexer_text_region)
|
||||
cropped_lines_meging_indexing.append(0)
|
||||
img_fin = np.ones((image_height, image_width, 3))*1
|
||||
cropped_lines.append(img_fin)
|
||||
|
||||
else:
|
||||
for indexing2, ind_poly in enumerate(ind_poly_first):
|
||||
cropped_lines_region_indexer.append(indexer_text_region)
|
||||
if not curved_line:
|
||||
ind_poly = copy.deepcopy(ind_poly)
|
||||
box_ind = all_box_coord[indexing]
|
||||
|
||||
ind_poly = return_textline_contour_with_added_box_coordinate(ind_poly, box_ind)
|
||||
#print(ind_poly_copy)
|
||||
ind_poly[ind_poly<0] = 0
|
||||
x, y, w, h = cv2.boundingRect(ind_poly)
|
||||
|
||||
w_scaled = w * image_height/float(h)
|
||||
|
||||
mask_poly = np.zeros(image.shape)
|
||||
|
||||
img_poly_on_img = np.copy(image)
|
||||
|
||||
mask_poly = cv2.fillPoly(mask_poly, pts=[ind_poly], color=(1, 1, 1))
|
||||
|
||||
|
||||
|
||||
mask_poly = mask_poly[y:y+h, x:x+w, :]
|
||||
img_crop = img_poly_on_img[y:y+h, x:x+w, :]
|
||||
|
||||
img_crop[mask_poly==0] = 255
|
||||
|
||||
if w_scaled < 640:#1.5*image_width:
|
||||
img_fin = preprocess_and_resize_image_for_ocrcnn_model(img_crop, image_height, image_width)
|
||||
cropped_lines.append(img_fin)
|
||||
cropped_lines_meging_indexing.append(0)
|
||||
else:
|
||||
splited_images, splited_images_bin = return_textlines_split_if_needed(img_crop, None)
|
||||
|
||||
if splited_images:
|
||||
img_fin = preprocess_and_resize_image_for_ocrcnn_model(splited_images[0],
|
||||
image_height,
|
||||
image_width)
|
||||
cropped_lines.append(img_fin)
|
||||
cropped_lines_meging_indexing.append(1)
|
||||
|
||||
img_fin = preprocess_and_resize_image_for_ocrcnn_model(splited_images[1],
|
||||
image_height,
|
||||
image_width)
|
||||
|
||||
cropped_lines.append(img_fin)
|
||||
cropped_lines_meging_indexing.append(-1)
|
||||
|
||||
else:
|
||||
img_fin = preprocess_and_resize_image_for_ocrcnn_model(img_crop,
|
||||
image_height,
|
||||
image_width)
|
||||
cropped_lines.append(img_fin)
|
||||
cropped_lines_meging_indexing.append(0)
|
||||
|
||||
indexer_text_region+=1
|
||||
|
||||
extracted_texts = []
|
||||
|
||||
n_iterations = math.ceil(len(cropped_lines) / b_s_ocr)
|
||||
|
||||
for i in range(n_iterations):
|
||||
if i==(n_iterations-1):
|
||||
n_start = i*b_s_ocr
|
||||
imgs = cropped_lines[n_start:]
|
||||
imgs = np.array(imgs)
|
||||
imgs = imgs.reshape(imgs.shape[0], image_height, image_width, 3)
|
||||
|
||||
|
||||
else:
|
||||
n_start = i*b_s_ocr
|
||||
n_end = (i+1)*b_s_ocr
|
||||
imgs = cropped_lines[n_start:n_end]
|
||||
imgs = np.array(imgs).reshape(b_s_ocr, image_height, image_width, 3)
|
||||
|
||||
|
||||
preds = prediction_model.predict(imgs, verbose=0)
|
||||
|
||||
pred_texts = decode_batch_predictions(preds, num_to_char)
|
||||
|
||||
for ib in range(imgs.shape[0]):
|
||||
pred_texts_ib = pred_texts[ib].replace("[UNK]", "")
|
||||
extracted_texts.append(pred_texts_ib)
|
||||
|
||||
extracted_texts_merged = [extracted_texts[ind]
|
||||
if cropped_lines_meging_indexing[ind]==0
|
||||
else extracted_texts[ind]+" "+extracted_texts[ind+1]
|
||||
if cropped_lines_meging_indexing[ind]==1
|
||||
else None
|
||||
for ind in range(len(cropped_lines_meging_indexing))]
|
||||
|
||||
extracted_texts_merged = [ind for ind in extracted_texts_merged if ind is not None]
|
||||
unique_cropped_lines_region_indexer = np.unique(cropped_lines_region_indexer)
|
||||
|
||||
ocr_all_textlines = []
|
||||
for ind in unique_cropped_lines_region_indexer:
|
||||
ocr_textline_in_textregion = []
|
||||
extracted_texts_merged_un = np.array(extracted_texts_merged)[np.array(cropped_lines_region_indexer)==ind]
|
||||
for it_ind, text_textline in enumerate(extracted_texts_merged_un):
|
||||
ocr_textline_in_textregion.append(text_textline)
|
||||
ocr_all_textlines.append(ocr_textline_in_textregion)
|
||||
return ocr_all_textlines
|
||||
|
|
@ -1,293 +0,0 @@
|
|||
# pylint: disable=too-many-locals,wrong-import-position,too-many-lines,too-many-statements,chained-comparison,fixme,broad-except,c-extension-no-member
|
||||
# pylint: disable=import-error
|
||||
from pathlib import Path
|
||||
import os.path
|
||||
import logging
|
||||
from typing import Optional
|
||||
import numpy as np
|
||||
from shapely import affinity, clip_by_rect
|
||||
|
||||
from ocrd_utils import points_from_polygon
|
||||
from ocrd_models.ocrd_page import (
|
||||
BorderType,
|
||||
CoordsType,
|
||||
TextLineType,
|
||||
TextEquivType,
|
||||
TextRegionType,
|
||||
ImageRegionType,
|
||||
TableRegionType,
|
||||
SeparatorRegionType,
|
||||
to_xml
|
||||
)
|
||||
|
||||
from .utils.xml import create_page_xml, xml_reading_order
|
||||
from .utils.counter import EynollahIdCounter
|
||||
from .utils.contour import contour2polygon, make_valid
|
||||
|
||||
class EynollahXmlWriter:
|
||||
|
||||
def __init__(self, *, dir_out, image_filename, image_width, image_height, curved_line, pcgts=None):
|
||||
self.logger = logging.getLogger('eynollah.writer')
|
||||
self.counter = EynollahIdCounter()
|
||||
self.dir_out = dir_out
|
||||
self.image_filename = image_filename
|
||||
self.output_filename = os.path.join(self.dir_out or "", self.image_filename_stem) + ".xml"
|
||||
self.curved_line = curved_line
|
||||
self.pcgts = pcgts
|
||||
self.image_height = image_height
|
||||
self.image_width = image_width
|
||||
self.scale_x = 1.0
|
||||
self.scale_y = 1.0
|
||||
|
||||
@property
|
||||
def image_filename_stem(self):
|
||||
return Path(Path(self.image_filename).name).stem
|
||||
|
||||
def calculate_points(self, contour, offset=None):
|
||||
poly = contour2polygon(contour)
|
||||
if offset is not None:
|
||||
poly = affinity.translate(poly, *offset)
|
||||
poly = affinity.scale(poly, xfact=1 / self.scale_x, yfact=1 / self.scale_y, origin=(0, 0))
|
||||
poly = make_valid(clip_by_rect(poly, 0, 0, self.image_width, self.image_height))
|
||||
return points_from_polygon(poly.exterior.coords[:-1])
|
||||
|
||||
def serialize_lines_in_region(self, text_region, all_found_textline_polygons, region_idx, page_coord, slopes, counter, ocr_all_textlines_textregion):
|
||||
for j, polygon_textline in enumerate(all_found_textline_polygons[region_idx]):
|
||||
coords = CoordsType()
|
||||
textline = TextLineType(id=counter.next_line_id, Coords=coords)
|
||||
if ocr_all_textlines_textregion:
|
||||
# FIXME: add OCR confidence
|
||||
textline.set_TextEquiv([TextEquivType(Unicode=ocr_all_textlines_textregion[j])])
|
||||
text_region.add_TextLine(textline)
|
||||
text_region.set_orientation(-slopes[region_idx])
|
||||
offset = [page_coord[2], page_coord[0]]
|
||||
coords.set_points(self.calculate_points(polygon_textline, offset))
|
||||
|
||||
def write_pagexml(self, pcgts):
|
||||
self.logger.info("output filename: '%s'", self.output_filename)
|
||||
with open(self.output_filename, 'w') as f:
|
||||
f.write(to_xml(pcgts))
|
||||
|
||||
def build_pagexml_no_full_layout(
|
||||
self,
|
||||
*,
|
||||
num_col,
|
||||
found_polygons_text_region,
|
||||
page_coord,
|
||||
page_slope,
|
||||
order_of_texts,
|
||||
all_found_textline_polygons,
|
||||
found_polygons_images,
|
||||
found_polygons_tables,
|
||||
found_polygons_marginals_left,
|
||||
found_polygons_marginals_right,
|
||||
all_found_textline_polygons_marginals_left,
|
||||
all_found_textline_polygons_marginals_right,
|
||||
slopes,
|
||||
slopes_marginals_left,
|
||||
slopes_marginals_right,
|
||||
cont_page,
|
||||
polygons_seplines,
|
||||
ocr_all_textlines=None,
|
||||
ocr_all_textlines_marginals_left=None,
|
||||
ocr_all_textlines_marginals_right=None,
|
||||
ocr_all_textlines_drop=None,
|
||||
conf_textregions=None,
|
||||
conf_marginals_left=None,
|
||||
conf_marginals_right=None,
|
||||
conf_images=None,
|
||||
conf_tables=None,
|
||||
):
|
||||
return self.build_pagexml_full_layout(
|
||||
num_col=num_col,
|
||||
found_polygons_text_region=found_polygons_text_region,
|
||||
found_polygons_text_region_h=[],
|
||||
page_coord=page_coord,
|
||||
page_slope=page_slope,
|
||||
order_of_texts=order_of_texts,
|
||||
all_found_textline_polygons=all_found_textline_polygons,
|
||||
all_found_textline_polygons_h=[],
|
||||
found_polygons_images=found_polygons_images,
|
||||
found_polygons_tables=found_polygons_tables,
|
||||
found_polygons_drop_capitals=[],
|
||||
found_polygons_marginals_left=found_polygons_marginals_left,
|
||||
found_polygons_marginals_right=found_polygons_marginals_right,
|
||||
all_found_textline_polygons_marginals_left=all_found_textline_polygons_marginals_left,
|
||||
all_found_textline_polygons_marginals_right=all_found_textline_polygons_marginals_right,
|
||||
slopes=slopes,
|
||||
slopes_h=[],
|
||||
slopes_marginals_left=slopes_marginals_left,
|
||||
slopes_marginals_right=slopes_marginals_right,
|
||||
cont_page=cont_page,
|
||||
polygons_seplines=polygons_seplines,
|
||||
ocr_all_textlines=ocr_all_textlines,
|
||||
ocr_all_textlines_marginals_left=ocr_all_textlines_marginals_left,
|
||||
ocr_all_textlines_marginals_right=ocr_all_textlines_marginals_right,
|
||||
conf_textregions=conf_textregions,
|
||||
conf_marginals_left=conf_marginals_left,
|
||||
conf_marginals_right=conf_marginals_right,
|
||||
conf_images=conf_images,
|
||||
conf_tables=conf_tables,
|
||||
)
|
||||
|
||||
def build_pagexml_full_layout(
|
||||
self,
|
||||
*,
|
||||
num_col,
|
||||
found_polygons_text_region,
|
||||
found_polygons_text_region_h,
|
||||
page_coord,
|
||||
page_slope,
|
||||
order_of_texts,
|
||||
all_found_textline_polygons,
|
||||
all_found_textline_polygons_h,
|
||||
found_polygons_images,
|
||||
found_polygons_tables,
|
||||
found_polygons_drop_capitals,
|
||||
found_polygons_marginals_left,
|
||||
found_polygons_marginals_right,
|
||||
all_found_textline_polygons_marginals_left,
|
||||
all_found_textline_polygons_marginals_right,
|
||||
slopes,
|
||||
slopes_h,
|
||||
slopes_marginals_left,
|
||||
slopes_marginals_right,
|
||||
cont_page,
|
||||
polygons_seplines,
|
||||
ocr_all_textlines=None,
|
||||
ocr_all_textlines_h=None,
|
||||
ocr_all_textlines_marginals_left=None,
|
||||
ocr_all_textlines_marginals_right=None,
|
||||
ocr_all_textlines_drop=None,
|
||||
conf_textregions=None,
|
||||
conf_textregions_h=None,
|
||||
conf_marginals_left=None,
|
||||
conf_marginals_right=None,
|
||||
conf_images=None,
|
||||
conf_tables=None,
|
||||
conf_drops=None,
|
||||
):
|
||||
self.logger.debug('enter build_pagexml')
|
||||
|
||||
# create the file structure
|
||||
pcgts = self.pcgts if self.pcgts else create_page_xml(
|
||||
self.image_filename, self.image_height, self.image_width)
|
||||
page = pcgts.get_Page()
|
||||
pcgts.Metadata.Comments = "num_col %d" % num_col
|
||||
page.set_custom('layout {num_col:%d;} ' % num_col)
|
||||
page.set_orientation(-page_slope)
|
||||
if len(cont_page):
|
||||
page.set_Border(BorderType(Coords=CoordsType(points=self.calculate_points(cont_page[0]))))
|
||||
|
||||
offset = [page_coord[2], page_coord[0]]
|
||||
counter = EynollahIdCounter()
|
||||
if len(order_of_texts):
|
||||
_counter_marginals = EynollahIdCounter(region_idx=len(order_of_texts))
|
||||
id_of_marginalia_left = [_counter_marginals.next_region_id
|
||||
for _ in found_polygons_marginals_left]
|
||||
id_of_marginalia_right = [_counter_marginals.next_region_id
|
||||
for _ in found_polygons_marginals_right]
|
||||
xml_reading_order(page, order_of_texts, id_of_marginalia_left, id_of_marginalia_right)
|
||||
|
||||
for mm, region_contour in enumerate(found_polygons_text_region):
|
||||
textregion = TextRegionType(
|
||||
id=counter.next_region_id, type_='paragraph',
|
||||
Coords=CoordsType(points=self.calculate_points(region_contour, offset))
|
||||
)
|
||||
assert textregion.Coords
|
||||
if conf_textregions:
|
||||
textregion.Coords.set_conf(conf_textregions[mm])
|
||||
page.add_TextRegion(textregion)
|
||||
if ocr_all_textlines:
|
||||
ocr_textlines = ocr_all_textlines[mm]
|
||||
else:
|
||||
ocr_textlines = None
|
||||
self.serialize_lines_in_region(textregion, all_found_textline_polygons, mm, page_coord,
|
||||
slopes, counter, ocr_textlines)
|
||||
|
||||
self.logger.debug('len(found_polygons_text_region_h) %s', len(found_polygons_text_region_h))
|
||||
for mm, region_contour in enumerate(found_polygons_text_region_h):
|
||||
textregion = TextRegionType(
|
||||
id=counter.next_region_id, type_='heading',
|
||||
Coords=CoordsType(points=self.calculate_points(region_contour, offset))
|
||||
)
|
||||
assert textregion.Coords
|
||||
if conf_textregions_h:
|
||||
textregion.Coords.set_conf(conf_textregions_h[mm])
|
||||
page.add_TextRegion(textregion)
|
||||
if ocr_all_textlines_h:
|
||||
ocr_textlines = ocr_all_textlines_h[mm]
|
||||
else:
|
||||
ocr_textlines = None
|
||||
self.serialize_lines_in_region(textregion, all_found_textline_polygons_h, mm, page_coord,
|
||||
slopes_h, counter, ocr_textlines)
|
||||
|
||||
for mm, region_contour in enumerate(found_polygons_drop_capitals):
|
||||
dropcapital = TextRegionType(
|
||||
id=counter.next_region_id, type_='drop-capital',
|
||||
Coords=CoordsType(points=self.calculate_points(region_contour, offset))
|
||||
)
|
||||
if conf_drops:
|
||||
dropcapital.Coords.set_conf(conf_drops[mm])
|
||||
page.add_TextRegion(dropcapital)
|
||||
slopes_drop = [0]
|
||||
if ocr_all_textlines_drop:
|
||||
ocr_textlines = ocr_all_textlines_drop[mm]
|
||||
else:
|
||||
ocr_textlines = None
|
||||
self.serialize_lines_in_region(dropcapital, [[found_polygons_drop_capitals[mm]]], 0, page_coord,
|
||||
slopes_drop, counter, ocr_textlines)
|
||||
|
||||
for mm, region_contour in enumerate(found_polygons_marginals_left):
|
||||
marginal = TextRegionType(
|
||||
id=counter.next_region_id, type_='marginalia',
|
||||
Coords=CoordsType(points=self.calculate_points(region_contour, offset))
|
||||
)
|
||||
if conf_marginals_left:
|
||||
marginal.Coords.set_conf(conf_marginals_left[mm])
|
||||
page.add_TextRegion(marginal)
|
||||
if ocr_all_textlines_marginals_left:
|
||||
ocr_textlines = ocr_all_textlines_marginals_left[mm]
|
||||
else:
|
||||
ocr_textlines = None
|
||||
self.serialize_lines_in_region(marginal, all_found_textline_polygons_marginals_left, mm, page_coord,
|
||||
slopes_marginals_left, counter, ocr_textlines)
|
||||
|
||||
for mm, region_contour in enumerate(found_polygons_marginals_right):
|
||||
marginal = TextRegionType(
|
||||
id=counter.next_region_id, type_='marginalia',
|
||||
Coords=CoordsType(points=self.calculate_points(region_contour, offset))
|
||||
)
|
||||
if conf_marginals_right:
|
||||
marginal.Coords.set_conf(conf_marginals_right[mm])
|
||||
page.add_TextRegion(marginal)
|
||||
if ocr_all_textlines_marginals_right:
|
||||
ocr_textlines = ocr_all_textlines_marginals_right[mm]
|
||||
else:
|
||||
ocr_textlines = None
|
||||
self.serialize_lines_in_region(marginal, all_found_textline_polygons_marginals_right, mm, page_coord,
|
||||
slopes_marginals_right, counter, ocr_textlines)
|
||||
|
||||
for mm, region_contour in enumerate(found_polygons_images):
|
||||
image = ImageRegionType(
|
||||
id=counter.next_region_id,
|
||||
Coords=CoordsType(points=self.calculate_points(region_contour, offset)))
|
||||
if conf_images:
|
||||
image.Coords.set_conf(conf_images[mm])
|
||||
page.add_ImageRegion(image)
|
||||
|
||||
for region_contour in polygons_seplines:
|
||||
page.add_SeparatorRegion(
|
||||
SeparatorRegionType(id=counter.next_region_id,
|
||||
Coords=CoordsType(points=self.calculate_points(region_contour, offset))))
|
||||
|
||||
for mm, region_contour in enumerate(found_polygons_tables):
|
||||
table = TableRegionType(
|
||||
id=counter.next_region_id,
|
||||
Coords=CoordsType(points=self.calculate_points(region_contour, offset)))
|
||||
if conf_tables:
|
||||
table.Coords.set_conf(conf_tables[mm])
|
||||
page.add_TableRegion(table)
|
||||
|
||||
return pcgts
|
||||
|
||||
54
tests/base.py
Normal file
54
tests/base.py
Normal file
|
|
@ -0,0 +1,54 @@
|
|||
# pylint: disable=unused-import
|
||||
|
||||
from os.path import dirname, realpath
|
||||
from os import chdir
|
||||
import sys
|
||||
import logging
|
||||
import io
|
||||
import collections
|
||||
from unittest import TestCase as VanillaTestCase, skip, main as unittests_main
|
||||
import pytest
|
||||
from ocrd_utils import disableLogging, initLogging
|
||||
|
||||
def main(fn=None):
|
||||
if fn:
|
||||
sys.exit(pytest.main([fn]))
|
||||
else:
|
||||
unittests_main()
|
||||
|
||||
class TestCase(VanillaTestCase):
|
||||
|
||||
@classmethod
|
||||
def setUpClass(cls):
|
||||
chdir(dirname(realpath(__file__)) + '/..')
|
||||
|
||||
def setUp(self):
|
||||
disableLogging()
|
||||
initLogging()
|
||||
|
||||
class CapturingTestCase(TestCase):
|
||||
"""
|
||||
A TestCase that needs to capture stderr/stdout and invoke click CLI.
|
||||
"""
|
||||
|
||||
@pytest.fixture(autouse=True)
|
||||
def _setup_pytest_capfd(self, capfd):
|
||||
self.capfd = capfd
|
||||
|
||||
def invoke_cli(self, cli, args):
|
||||
"""
|
||||
Substitution for click.CliRunner.invooke that works together nicely
|
||||
with unittests/pytest capturing stdout/stderr.
|
||||
"""
|
||||
self.capture_out_err() # XXX snapshot just before executing the CLI
|
||||
code = 0
|
||||
sys.argv[1:] = args # XXX necessary because sys.argv reflects pytest args not cli args
|
||||
try:
|
||||
cli.main(args=args)
|
||||
except SystemExit as e:
|
||||
code = e.code
|
||||
out, err = self.capture_out_err()
|
||||
return code, out, err
|
||||
|
||||
def capture_out_err(self):
|
||||
return self.capfd.readouterr()
|
||||
|
|
@ -1,47 +0,0 @@
|
|||
from typing import List
|
||||
import pytest
|
||||
import logging
|
||||
|
||||
from click.testing import CliRunner, Result
|
||||
from eynollah.cli import main as eynollah_cli
|
||||
|
||||
|
||||
@pytest.fixture
|
||||
def run_eynollah_ok_and_check_logs(
|
||||
pytestconfig,
|
||||
caplog,
|
||||
model_dir,
|
||||
eynollah_subcommands,
|
||||
eynollah_log_filter,
|
||||
):
|
||||
"""
|
||||
Generates a Click Runner for `cli`, injects model_path and logging level
|
||||
to `args`, runs the command and checks whether the logs generated contain
|
||||
every fragment in `expected_logs`
|
||||
"""
|
||||
|
||||
def _run_click_ok_logs(
|
||||
subcommand: 'str',
|
||||
args: List[str],
|
||||
expected_logs: List[str],
|
||||
) -> Result:
|
||||
assert subcommand in eynollah_subcommands, f'subcommand {subcommand} must be one of {eynollah_subcommands}'
|
||||
args = [
|
||||
'-m', model_dir,
|
||||
subcommand,
|
||||
*args
|
||||
]
|
||||
if pytestconfig.getoption('verbose') > 0:
|
||||
args = ['-l', 'DEBUG'] + args
|
||||
caplog.set_level(logging.INFO)
|
||||
runner = CliRunner()
|
||||
with caplog.filtering(eynollah_log_filter):
|
||||
result = runner.invoke(eynollah_cli, args, catch_exceptions=False)
|
||||
assert result.exit_code == 0, result.stdout
|
||||
if expected_logs:
|
||||
logmsgs = [logrec.message for logrec in caplog.records]
|
||||
assert any(logmsg.startswith(needle) for needle in expected_logs for logmsg in logmsgs), f'{expected_logs} not in {logmsgs}'
|
||||
return result
|
||||
|
||||
return _run_click_ok_logs
|
||||
|
||||
|
|
@ -1,53 +0,0 @@
|
|||
import pytest
|
||||
from PIL import Image
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
"options",
|
||||
[
|
||||
[], # defaults
|
||||
["--no-patches"],
|
||||
], ids=str)
|
||||
def test_run_eynollah_binarization_filename(
|
||||
tmp_path,
|
||||
run_eynollah_ok_and_check_logs,
|
||||
resources_dir,
|
||||
options,
|
||||
):
|
||||
infile = resources_dir / '2files/kant_aufklaerung_1784_0020.tif'
|
||||
outfile = tmp_path / 'kant_aufklaerung_1784_0020.png'
|
||||
run_eynollah_ok_and_check_logs(
|
||||
'binarization',
|
||||
[
|
||||
'-i', str(infile),
|
||||
'-o', str(outfile),
|
||||
] + options,
|
||||
[
|
||||
f"output filename: '{str(outfile)}'"
|
||||
]
|
||||
)
|
||||
assert outfile.exists()
|
||||
with Image.open(infile) as original_img:
|
||||
original_size = original_img.size
|
||||
with Image.open(outfile) as binarized_img:
|
||||
binarized_size = binarized_img.size
|
||||
assert original_size == binarized_size
|
||||
|
||||
def test_run_eynollah_binarization_directory(
|
||||
tmp_path,
|
||||
run_eynollah_ok_and_check_logs,
|
||||
resources_dir,
|
||||
image_resources,
|
||||
):
|
||||
outdir = tmp_path
|
||||
run_eynollah_ok_and_check_logs(
|
||||
'binarization',
|
||||
[
|
||||
'-di', str(resources_dir / '2files'),
|
||||
'-o', str(outdir),
|
||||
],
|
||||
[
|
||||
str(image_resources[0]),
|
||||
str(image_resources[1]),
|
||||
]
|
||||
)
|
||||
assert len(list(outdir.iterdir())) == 2
|
||||
|
|
@ -1,55 +0,0 @@
|
|||
import pytest
|
||||
from PIL import Image
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
"options",
|
||||
[
|
||||
[], # defaults
|
||||
["-sos"],
|
||||
], ids=str)
|
||||
def test_run_eynollah_enhancement_filename(
|
||||
tmp_path,
|
||||
resources_dir,
|
||||
run_eynollah_ok_and_check_logs,
|
||||
options,
|
||||
):
|
||||
infile = resources_dir / '2files/kant_aufklaerung_1784_0020.tif'
|
||||
outfile = tmp_path / 'kant_aufklaerung_1784_0020.png'
|
||||
run_eynollah_ok_and_check_logs(
|
||||
'enhancement',
|
||||
[
|
||||
'-i', str(infile),
|
||||
'-o', str(outfile.parent),
|
||||
# force rescaling
|
||||
'-ncu', 3,
|
||||
] + options,
|
||||
[
|
||||
'Enhancement applied',
|
||||
]
|
||||
)
|
||||
with Image.open(infile) as original_img:
|
||||
original_size = original_img.size
|
||||
with Image.open(outfile) as enhanced_img:
|
||||
enhanced_size = enhanced_img.size
|
||||
assert (original_size == enhanced_size) == ("-sos" in options)
|
||||
|
||||
def test_run_eynollah_enhancement_directory(
|
||||
tmp_path,
|
||||
resources_dir,
|
||||
image_resources,
|
||||
run_eynollah_ok_and_check_logs,
|
||||
):
|
||||
outdir = tmp_path
|
||||
run_eynollah_ok_and_check_logs(
|
||||
'enhancement',
|
||||
[
|
||||
'-di', str(resources_dir/ '2files'),
|
||||
'-o', str(outdir),
|
||||
# force rescaling
|
||||
'-ncu', 3,
|
||||
],
|
||||
[
|
||||
'Enhancement applied',
|
||||
]
|
||||
)
|
||||
assert len(list(outdir.iterdir())) == 2
|
||||
|
|
@ -1,119 +0,0 @@
|
|||
import pytest
|
||||
from ocrd_modelfactory import page_from_file
|
||||
from ocrd_models.constants import NAMESPACES as NS
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
"options",
|
||||
[
|
||||
[], # defaults
|
||||
#["--allow_scaling", "--curved-line"],
|
||||
["--allow_scaling", "--curved-line", "--full-layout"],
|
||||
["--allow_scaling", "--curved-line", "--full-layout", "--reading_order_machine_based"],
|
||||
# -ep ...
|
||||
# -eoi ...
|
||||
# --skip_layout_and_reading_order
|
||||
], ids=str)
|
||||
def test_run_eynollah_layout_filename(
|
||||
tmp_path,
|
||||
run_eynollah_ok_and_check_logs,
|
||||
resources_dir,
|
||||
options,
|
||||
):
|
||||
infile = resources_dir / '2files/kant_aufklaerung_1784_0020.tif'
|
||||
outfile = tmp_path / 'kant_aufklaerung_1784_0020.xml'
|
||||
run_eynollah_ok_and_check_logs(
|
||||
'layout',
|
||||
[
|
||||
'-i', str(infile),
|
||||
'-o', str(outfile.parent),
|
||||
] + options,
|
||||
[
|
||||
str(infile)
|
||||
]
|
||||
)
|
||||
assert outfile.exists()
|
||||
tree = page_from_file(str(outfile)).etree
|
||||
regions = tree.xpath("//page:TextRegion", namespaces=NS)
|
||||
assert len(regions) >= 2, "result is inaccurate"
|
||||
regions = tree.xpath("//page:SeparatorRegion", namespaces=NS)
|
||||
assert len(regions) >= 2, "result is inaccurate"
|
||||
lines = tree.xpath("//page:TextLine", namespaces=NS)
|
||||
assert len(lines) == 31, "result is inaccurate" # 29 paragraph lines, 1 page and 1 catch-word line
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
"options",
|
||||
[
|
||||
["--tables"],
|
||||
["--tables", "--full-layout"],
|
||||
], ids=str)
|
||||
def test_run_eynollah_layout_filename2(
|
||||
tmp_path,
|
||||
resources_dir,
|
||||
run_eynollah_ok_and_check_logs,
|
||||
options,
|
||||
):
|
||||
infile = resources_dir / '2files/euler_rechenkunst01_1738_0025.tif'
|
||||
outfile = tmp_path / 'euler_rechenkunst01_1738_0025.xml'
|
||||
run_eynollah_ok_and_check_logs(
|
||||
'layout',
|
||||
[
|
||||
'-i', str(infile),
|
||||
'-o', str(outfile.parent),
|
||||
] + options,
|
||||
[
|
||||
str(infile)
|
||||
]
|
||||
)
|
||||
assert outfile.exists()
|
||||
tree = page_from_file(str(outfile)).etree
|
||||
regions = tree.xpath("//page:TextRegion", namespaces=NS)
|
||||
assert len(regions) >= 2, "result is inaccurate"
|
||||
regions = tree.xpath("//page:TableRegion", namespaces=NS)
|
||||
# model/decoding is not very precise, so (depending on mode) we can get fractures/splits/FP
|
||||
assert len(regions) >= 1, "result is inaccurate"
|
||||
regions = tree.xpath("//page:SeparatorRegion", namespaces=NS)
|
||||
assert len(regions) >= 2, "result is inaccurate"
|
||||
lines = tree.xpath("//page:TextLine", namespaces=NS)
|
||||
assert len(lines) >= 2, "result is inaccurate" # mostly table (if detected correctly), but 1 page and 1 catch-word line
|
||||
|
||||
def test_run_eynollah_layout_directory(
|
||||
tmp_path,
|
||||
resources_dir,
|
||||
run_eynollah_ok_and_check_logs,
|
||||
):
|
||||
outdir = tmp_path
|
||||
run_eynollah_ok_and_check_logs(
|
||||
'layout',
|
||||
[
|
||||
'-di', str(resources_dir / '2files'),
|
||||
'-o', str(outdir),
|
||||
],
|
||||
[
|
||||
'Job done in',
|
||||
'All jobs done in',
|
||||
]
|
||||
)
|
||||
assert len(list(outdir.iterdir())) == 2
|
||||
|
||||
# def test_run_eynollah_layout_marginalia(
|
||||
# tmp_path,
|
||||
# resources_dir,
|
||||
# run_eynollah_ok_and_check_logs,
|
||||
# ):
|
||||
# outdir = tmp_path
|
||||
# outfile = outdir / 'estor_rechtsgelehrsamkeit02_1758_0880_800px.xml'
|
||||
# run_eynollah_ok_and_check_logs(
|
||||
# 'layout',
|
||||
# [
|
||||
# '-i', str(resources_dir / 'estor_rechtsgelehrsamkeit02_1758_0880_800px.jpg'),
|
||||
# '-o', str(outdir),
|
||||
# ],
|
||||
# [
|
||||
# 'Job done in',
|
||||
# 'All jobs done in',
|
||||
# ]
|
||||
# )
|
||||
# assert outfile.exists()
|
||||
# tree = page_from_file(str(outfile)).etree
|
||||
# regions = tree.xpath('//page:TextRegion[type="marginalia"]', namespaces=NS)
|
||||
# assert len(regions) == 5, "expected 5 marginalia regions"
|
||||
|
|
@ -1,47 +0,0 @@
|
|||
from ocrd_modelfactory import page_from_file
|
||||
from ocrd_models.constants import NAMESPACES as NS
|
||||
|
||||
def test_run_eynollah_mbreorder_filename(
|
||||
tmp_path,
|
||||
resources_dir,
|
||||
run_eynollah_ok_and_check_logs,
|
||||
):
|
||||
infile = resources_dir / '2files/kant_aufklaerung_1784_0020.xml'
|
||||
outfile = tmp_path /'kant_aufklaerung_1784_0020.xml'
|
||||
run_eynollah_ok_and_check_logs(
|
||||
'machine-based-reading-order',
|
||||
[
|
||||
'-i', str(infile),
|
||||
'-o', str(outfile.parent),
|
||||
],
|
||||
[
|
||||
# FIXME: mbreorder has no logging!
|
||||
]
|
||||
)
|
||||
assert outfile.exists()
|
||||
#in_tree = page_from_file(str(infile)).etree
|
||||
#in_order = in_tree.xpath("//page:OrderedGroup//@regionRef", namespaces=NS)
|
||||
out_tree = page_from_file(str(outfile)).etree
|
||||
out_order = out_tree.xpath("//page:OrderedGroup//@regionRef", namespaces=NS)
|
||||
#assert len(out_order) >= 2, "result is inaccurate"
|
||||
#assert in_order != out_order
|
||||
assert out_order == ['r_1_1', 'r_2_1', 'r_2_2', 'r_2_3']
|
||||
|
||||
def test_run_eynollah_mbreorder_directory(
|
||||
tmp_path,
|
||||
resources_dir,
|
||||
run_eynollah_ok_and_check_logs,
|
||||
):
|
||||
outdir = tmp_path
|
||||
run_eynollah_ok_and_check_logs(
|
||||
'machine-based-reading-order',
|
||||
[
|
||||
'-di', str(resources_dir / '2files'),
|
||||
'-o', str(outdir),
|
||||
],
|
||||
[
|
||||
# FIXME: mbreorder has no logging!
|
||||
]
|
||||
)
|
||||
assert len(list(outdir.iterdir())) == 2
|
||||
|
||||
Some files were not shown because too many files have changed in this diff Show more
Loading…
Add table
Add a link
Reference in a new issue