- layout/extract-images: when running with `input_binary`, then
pass the binarized image of the full page to the writer
- writer: if there is a binarized image, then add an `AlternativeImage`
to the top-level of the output PAGE (with the array instead of a
file name)
- writer.write_xml (standalone CLIs): save the image in the same
directory as the output file, but with `.bin.png` isntead of
`.xml` as suffix
- processor: convert the binarized image array to PIL and pass over
as OcrdPageResultImage for the final path name
- `read_xml()`: bring sanity here…
- use identifiers for class labels instead of numeric literals
- use functions from `ocrd_utils` for PAGE coordinates
- use PrintSpace/Border coordinates directly instead of plotting
and contour detection
- simplify ad-hoc PAGE XML parser and writer (a lot)
- avoid XML invalidity when `AlternativeImage` exists
- separate `run_single()` from `run()` (as in superclass)
- re-use superclass' `do_order_of_regions_with_model()`
- include drop capitals, too (as in superclass)
- CLI: add `--overwrite`, too
- fix model loading for predictor
- add `extract_images` model to ModelZoo and Makefile, too
- rename `get_regions_light_v_extract_only_images` → `get_early_layout`
(with similar refactoring as done in superclass previously)
- separate `run_single()` from `run()` (as in superclass)
- adapt to new `Region` dataclasses, too
- `utils`:
- add dataclasses `Region` (including contour, skew, conf,
centers and area – the latter computed during init, but only once
instead of over and over) and `TextRegion` (ditto, but with a list
of lines) to encapsulate intermediate results more succinctly
- drop `check_any_text_region_in_model_one_is_main_or_header`
- `run_single()`:
- instantiate `Region` and `TextRegion` lists for separator lines,
drop-caps, marginals, paragraphs, headings, images, tables instead
of multiple sets of identifiers for these segment types, name them
more consistently
- apply contour "dilation" (really Shapely polygon instantiation,
repair/join and buffering) only once in the writer after rescaling
to original/final image size
- `filter_contours_inside_a_bigger_one()`: separate and rename
- → `filter_small_regions()`: taking lists of `Region` (original and
deskewed) for paragraphs and marginals, filtering small regions
that are contained in large regions or marginals in-place
- → `filter_small_textlines()`: taking a list of `TextRegion` and
filtering lines that are contained in larger lines (of possibly
other regions) in-place
- `filter_contours_without_textlines_inside()`: simplify and rename
→ `filter_textregions_without_textlines()`
- rename `separate_marginals_to_left_and_right_and_order_from_top_to_down`
→ `separate_marginals_and_order()` and simplify
- `get_slopes_and_deskew*`: work on list of `Region` in-place
(modifying their `.skew` and `.lines`), extract line confidence, too
- writer:
- expect `Region` and `TextRegion` instances for various segment types
instead of wild list of identifiers
- simplify (a lot)
- merge `build_pagexml_no_full_layout` and build_pagexml_full_layout`
→ `build_pagexml()` (w/ empty lists as default)
- also annotate textline confidence
- enlarge/dilate/buffer coordinates differently for various segment types:
6px for text and tables, 2px for images and seplines
- add more type hints
- add CLI options `--num-jobs` and `--halt-fail`
- separate `Eynollah_ocr.run_single()` to be scheduled
- use ProcessPoolExecutor w/ forking and QueueListener
- also skip if input XML file is missing
- log processing times per job and overall
- training.models.CTCDecoder: prefer beam search over greedy
(because it is more accurate)
- training reload makefile: skip ONNX and TF-Serving conversion
for cnn-rnn-ocr models (because these would not work)
- training reload makefile: default to onnx and tf conversions
for all models (tf for training and onnx for inference)
instead of tf-serving export
