Robert Sachunsky
a1904fa660
tests: cover layout with OCR in various modes
2025-10-09 20:14:11 +02:00
Robert Sachunsky
595ed02743
run_single: simplify; allow running TrOCR in non-fl mode, too
...
- refactor final `self.full_layout` conditional, removing copied code
- allow running `self.ocr` and `self.tr` branch in both cases (non/fl)
- when running TrOCR, use model / processor / device initialised during init
(instead of ad-hoc loading)
2025-10-09 20:14:11 +02:00
Robert Sachunsky
6e57ab3741
textline_contours_postprocessing: do not catch arbitrary exceptions
2025-10-09 20:14:11 +02:00
Robert Sachunsky
fe603188f4
avoid unnecessary 3-channel conversions
2025-10-09 20:14:11 +02:00
Robert Sachunsky
155b8f68b8
matching deskewed text region contours with predicted: improve
...
- avoid duplicate and missing mappings by using a different approach:
instead of just minimising the center distance for the N contours
that we expect,
1. get all N:M distances
2. iterate over them from small to large
3. continue adding correspondences until both every original contour
and every deskewed contour have at least one match
4. where one original matches multiple deskewed contours,
join the latter polygons to map as single contour
5. where one deskewed contour matches multiple originals,
split the former by intersecting with each of the latter
(after bringing them into the same coordinate space),
so ultimately only the respective match gets assigned
2025-10-09 20:14:11 +02:00
Robert Sachunsky
0e00d7868b
matching deskewed text region contours with predicted: improve
...
- apply same min-area filter to deskewed contours as to original ones
2025-10-09 20:14:11 +02:00
Robert Sachunsky
0f33c21eb3
matching deskewed text region contours with predicted: improve
...
- when matching undeskewed and new contours, do not just
pick the closest centers, respectively, but also of similar
size (by making the contour area the 3rd dimension of the
vector norm in the distance calculation)
2025-10-09 20:14:11 +02:00
Robert Sachunsky
73e5a1def8
matching deskewed text region contours with predicted: simplify
...
- (no need for argmax if already sorted)
2025-10-09 20:14:11 +02:00
Robert Sachunsky
d774a23daa
matching deskewed text region contours with predicted: simplify
...
- avoid loops in favour of array processing
- improve readability and identifiers
2025-10-09 20:14:11 +02:00
Robert Sachunsky
29b4527bde
do_order_of_regions: simplify
...
- remove duplicate code via inline def for the try-catch
2025-10-09 20:14:11 +02:00
Robert Sachunsky
e674ea08f3
do_order_of_regions: drop redundant no/full_layout
...
(`_no_full_layout` is the same copied code as `_full_layout`;
the latter runs just the same if passed an empty list for headings)
2025-10-09 20:14:11 +02:00
Robert Sachunsky
e9bb62bd86
do_order_of_regions: simplify
...
- avoid loops in favour of array processing
2025-10-09 20:14:11 +02:00
Robert Sachunsky
7387f5a929
do_order_of_regions: improve box matching, simplify
...
- when searching for boxes matching contour, be more precise:
- avoid heuristic rules ("xmin + 80 within xrange") in favour
of exact criteria (contour properly contained in box)
- for fallback criterion (nearest centers), also require
proper containment of center in box
- `order_of_regions`: remove (now) unnecessary (and insufficient)
workaround for missing indexes (if boxes are not covering contours
exactly)
2025-10-09 20:14:11 +02:00
Robert Sachunsky
4950e6bd78
order_of_regions: simplify
...
- use new `find_center_of_contours`
- avoid unused calculations
- avoid loops in favour of array processing
2025-10-09 20:14:10 +02:00
Robert Sachunsky
a1c8fd4467
do_order_of_regions / order_of_regions: simplify
...
- array-convert only once (before returning from `order_of_regions`)
- avoid passing `matrix_of_orders` unnecessarily between
`order_of_regions` and `order_and_id_of_texts`
2025-10-09 20:14:10 +02:00
Robert Sachunsky
415b2cbad8
eynollah, drop_capitals: simplify
...
- use new `find_center_of_contours`
2025-10-09 20:14:10 +02:00
Robert Sachunsky
3f3353ec3a
do_order_of_regions: simplify
...
- avoid loops in favour of array processing
2025-10-09 20:14:10 +02:00
Robert Sachunsky
8c3d5eb0eb
separate_marginals_to_left_and_right_and_order_from_top_to_down: simplify
...
- use new `find_center_of_contours`
- avoid loops in favour of array processing
- avoid repeated sorting
2025-10-09 20:14:10 +02:00
kba
8215814a3f
Merge branch 'changelog-v0.5.0'
2025-10-09 14:03:45 +02:00
kba
4ffe6190d2
📝 changelog
2025-10-09 14:03:26 +02:00
vahidrezanezhad
8869c20c33
updating CHANGELOG for v0.5.0
2025-10-09 13:54:29 +02:00
Robert Sachunsky
81827c2942
filter_contours_inside_a_bigger_one: simplify
...
- use new `find_center_of_contours`
- avoid loops in favour of array processing
- use sets instead of `np.unique` and `np.delete` instead of list.pop
2025-10-06 13:32:34 +02:00
Robert Sachunsky
0b9d4901a6
contour features: avoid unused calculations, simplify, add shortcuts
...
- new function: `find_center_of_contours`
- simplified: `find_(new_)features_of_contours`
2025-10-02 20:51:03 +02:00
kba
8a9b4f8f55
remove commented-out requirement for tf == 2.12.1, rely on same version as in eynollah proper
2025-10-02 12:16:26 +02:00
kba
f60e0543ab
training: update docs
2025-10-01 19:16:58 +02:00
kba
1c043c586a
eynollah-training: all training CLI into single click group
2025-10-01 19:16:45 +02:00
kba
690d47444c
make relative wildcard imports explicit
2025-10-01 18:43:20 +02:00
kba
2baf42e878
organize imports, use relative imports
2025-10-01 18:15:54 +02:00
kba
4f5cdf3140
move training scripts to src/eynollah/training
2025-10-01 18:12:45 +02:00
kba
f0ef2b5db2
remove unused imports
2025-10-01 18:10:13 +02:00
kba
95bb5908bb
Merge branch 'integrate-training-from-sbb_pixelwise_segmentation' of https://github.com/qurator-spk/eynollah into integrate-training-from-sbb_pixelwise_segmentation
2025-10-01 18:02:09 +02:00
kba
48266b1ee0
make training dependencies optional-dependencies of eynollah
...
i.e. `pip install "eynollah[training]"` will install the requirements for training
2025-10-01 18:01:25 +02:00
kba
733af1e9a7
📝 update train/README.md, align with docs/train.md
2025-10-01 17:43:32 +02:00
vahidrezanezhad
5725e4fd1f
-Continue processing when num_col is None but textregions exist. -Convert marginal-only to main body if no main body is present. -Reset deskew angle to 0 when text region density (textregion area to page area) < 0.3 and angle > 45°.
2025-10-01 15:58:03 +02:00
cneud
4514d417a7
force GH markdown code block in list
2025-10-01 01:16:25 +02:00
cneud
e027bc038e
Update README.md
2025-10-01 01:05:15 +02:00
cneud
91d2a74ac9
remove redundant parentheses
2025-10-01 00:38:01 +02:00
cneud
f2f93e0251
list literal is faster than using list constructor to create a new list
2025-10-01 00:26:27 +02:00
cneud
70af00182b
mutable defaults are the source of all evil
2025-10-01 00:20:18 +02:00
cneud
1d0616eb69
comparisons to None should not use the equality operators
2025-10-01 00:15:11 +02:00
cneud
9ce127eb51
remove unnecessary backslash
2025-10-01 00:04:53 +02:00
cneud
558867eb24
fix typo
2025-10-01 00:04:07 +02:00
Robert Sachunsky
3aa7ad04fa
📝 update changelog
2025-09-30 23:14:52 +02:00
Robert Sachunsky
f0de1adabf
rm loky dependency
2025-09-30 23:12:18 +02:00
Robert Sachunsky
7daec392b9
Dockerfile: fix up CUDA installation for mixed TF/Torch
2025-09-30 22:10:45 +02:00
Robert Sachunsky
ad129ed46c
CI: remove OS from model cache keys
2025-09-30 22:05:53 +02:00
Robert Sachunsky
c86e59f481
CI: update model key, split up cache restore/save
2025-09-30 22:03:46 +02:00
Robert Sachunsky
a3d8197930
makefile: update model URL
2025-09-30 21:50:21 +02:00
Robert Sachunsky
61b20cc83d
tests: switch from subtests to parametrize, use --isolate everywhere to free CUDA memory in between
2025-09-30 19:20:35 +02:00
Robert Sachunsky
375e0263d4
CNN-RNN OCR model: switch to 20250930 version (compatible with TF 2.12 on CPU as well)
2025-09-30 19:16:50 +02:00
