You cannot select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

3.0 KiB

Raw Blame History

My OCR-D workflow

WIP. Given a OCR-D workspace with document images in the OCR-D-IMG file group, this workflow produces:

Binarized images
Line segmentation
OCR text (using Calamari and Tesseract, both with GT4HistOCR models)
(Given ground truth in OCR-D-GT-PAGE, also an OCR text evaluation report)

If you're interested in the exact processors, versions and parameters, please take a look at the script and possibly the individual Dockerfiles.

Goal

Provide a test environment to produce OCR output for historical prints, using OCR-D, especially ocrd_calamari and sbb_textline_detection, including all dependencies in Docker.

How to use

It's easiest to use it as pre-built containers. To run the containers on an example workspace:

# Update to the latest stable containers
~/devel/my_ocrd_workflow/run-docker-hub-update

# Download an example workspace
cd /tmp
wget https://qurator-data.de/examples/actevedef_718448162.first-page.zip
unzip actevedef_718448162.first-page.zip

# Run the workflow on it
cd actevedef_718448162.first-page
~/devel/my_ocrd_workflow/run-docker-hub

Build the containers yourself

To build the containers yourself using Docker:

cd ~/devel/my_ocrd_workflow
./build

You may then use the script run to use your self-built containers, analogous to the example above.

Viewing results

You may then examine the results using PRImA's PAGE Viewer:

java -jar /path/to/JPageViewer.jar \
  --resolve-dir . \
  OCR-D-OCR-CALAMARI/OCR-D-OCR-CALAMARI_00000024.xml

The workflow also produces OCR evaluation reports using dinglehopper, if ground truth was available:

firefox OCR-D-OCR-CALAMARI-EVAL/OCR-D-OCR-CALAMARI-EVAL_00000024.html

ppn2ocr

The ppn2ocr script produces a METS file with the best images for a given document in the State Library Berlin (SBB)'s digitized collection.

Install it with an up-to-date pip (otherwise this will fail due to a opencv-python-headless build failure):

pip install -r ~/devel/my_ocrd_workflow/requirements-ppn2ocr.txt

The document must be specified by its PPN, for example:

~/devel/my_ocrd_workflow/ppn2ocr PPN77164308X
cd PPN77164308X
~/devel/my_ocrd_workflow/run-docker-hub -I BEST --skip-validation

This produces a workspace directory PPN77164308X with the OCR results in it; the results are viewable as explained above.

ppn2ocr requires a working Docker setup and properly set up environment variables for the proxy configuration. At SBB, please read howto/docker-proxy.md and howto/proxy-settings-for-shell+python.md (in qurator's mono-repo).

3.0 KiB Raw Blame History