Feat/improve depth entry recognition

[redur]

Tune some parameters to improve depth entry recognition.
Introduce new checks:

• check if depths are strictly increasing.
• check if correcting a common OCR mistake makes a depth column valid.

Overall, more depth entries are now recognized. For the zurich dataset, even the layer detection is improved. For the geoquad dataset, the layer detection rate is approximately the same.

[redur]

This is just fixing a bug where we only kept rows with coordinates in the document_level_metrics

[redur]

This simply improved the scores.

[stijnvermeeren-swisstopo]

Keeping the segment length at 6 and simplifying the threshold to abs(scale_pearson_correlation_coef) >= 0.9999 regardless of segment length seems to work even better. (Implemented in my commit.)

[github-actions]

Coverage Report

File	Stmts	Miss	Cover	Missing
src/stratigraphy
__init__.py	8	1	88%	11
extract.py	188	188	0%	3–482
get_files.py	21	21	0%	3–48
line_detection.py	26	26	0%	3–76
main.py	94	94	0%	3–237
src/stratigraphy/util
boundarydepthcolumnvalidator.py	37	16	57%	47, 57, 60, 81–84, 102–118, 130–132
coordinate_extraction.py	127	7	94%	31, 62, 75–76, 80, 205, 328
dataclasses.py	32	3	91%	37–39
depthcolumn.py	192	64	67%	26, 30, 51, 57, 60–61, 85, 88, 95, 102, 110–111, 121, 138–154, 192, 229, 245–253, 264, 269, 276, 307, 312–319, 334–335, 378–420
depthcolumnentry.py	26	7	73%	12, 15, 29–30, 33, 45, 52
description_block_splitter.py	70	2	97%	24, 139
draw.py	80	80	0%	3–244
duplicate_detection.py	51	51	0%	3–146
extract_text.py	27	2	93%	38–39
find_depth_columns.py	91	6	93%	42–43, 71, 83, 176–177
find_description.py	63	28	56%	27–35, 50–63, 79–95, 172–175
geometric_line_utilities.py	86	2	98%	82, 132
interval.py	106	55	48%	24–27, 31–34, 39, 44, 47, 57–59, 99–145, 166, 171–187
language_detection.py	18	18	0%	3–45
layer_identifier_column.py	91	91	0%	3–227
line.py	49	4	92%	25, 42, 51, 98
linesquadtree.py	46	1	98%	76
plot_utils.py	43	43	0%	3–120
predictions.py	187	187	0%	3–387
textblock.py	74	8	89%	27, 51, 63, 75, 98, 119, 127, 155
util.py	40	18	55%	22, 40–47, 61–63, 87–88, 100–105
TOTAL	1873	1023	45%

Tests	Skipped	Failures	Errors	Time
61	0 💤	0 ❌	0 🔥	0.952s ⏱️

[redur]

Admittably, this is probably overfitted. Will be interesting to see the real scores ones we have a test set.

[stijnvermeeren-swisstopo]

Thanks for the work!
The code for correcting OCR mistakes does not make a big difference yet, but offers more potential for further improvements in the future.

With the additional changes that I've implemented (which I believe make the logic slightly simpler as well), I now get the following numbers (overall F1 score):

Zurich dataset:

• main branch: 87.16%
• Renato's version: 87.91%
• Stijn's version: 87.90%

Geoquat validation dataset:

• main branch: 49.60%
• Renato's version: 49.43%
• Stijn's version: 49.53%

While I believe that some further optimization of the parameters involved might be possible, the improvements will be marginal, and probably not worth the effort. Therefore, I would propose to merge the PR like this.