qurator-spk/dinglehopper
1
0
Fork 0
mirror of https://github.com/qurator-spk/dinglehopper.git synced 2025-06-09 11:50:00 +02:00
Code Issues Releases Wiki Activity

Reduce number of splits for short (one char) elements

Browse source
This commit is contained in:
Benjamin Rosemann 2020-11-13 11:45:55 +01:00
parent c9219cbacd
commit 0ef7810dd0
2 changed files with 9 additions and 5 deletions
Download patch file Download diff file Expand all files Collapse all files

10
qurator/dinglehopper/flexible_character_accuracy.py
View file

@ -145,9 +145,11 @@ def match_longest_gt_lines(
score = 0 if not match else character_accuracy(match.dist)
if score > best_score:
best_score, best_match, best_gt, best_ocr = score, match, gt_line, ocr_line
# early breaking: we only need one perfect fit
if best_score >= 1:
break
# Step 4 of the flexible character accuracy algorithm.
# Remove on full match or split.
if best_match:
remove_or_split(best_gt, best_match.gt, gt_lines)
remove_or_split(best_ocr, best_match.ocr, ocr_lines)
@ -168,7 +170,7 @@ def match_gt_line(
"""
min_penalty = float("inf")
best_match, best_ocr = None, None
for ocr_line in [*ocr_lines]:
for ocr_line in ocr_lines:
match = match_lines(gt_line, ocr_line)
if match:
penalty = calculate_penalty(gt_line, ocr_line, match, coef)
@ -233,7 +235,7 @@ def match_lines(gt_line: "Part", ocr_line: "Part") -> Optional[Match]:
for j, ocr_part in ocr_parts:
match = distance(gt_part, ocr_part)
edit_dist = score_edit_distance(match)
if edit_dist < min_edit_dist:
if edit_dist < min_edit_dist and match.dist.replace < min_length:
min_edit_dist = edit_dist
best_match = match
best_i, best_j = i, j
@ -247,7 +249,7 @@ def match_lines(gt_line: "Part", ocr_line: "Part") -> Optional[Match]:
ocr_line.substring(rel_start=best_j, rel_end=best_j + k),
)
edit_dist = score_edit_distance(match)
if edit_dist < min_edit_dist:
if edit_dist < min_edit_dist and match.dist.replace < min_length:
min_edit_dist = edit_dist
best_match = match
# is delete a better option?

4
qurator/dinglehopper/tests/test_flexible_character_accuracy.py
View file

@ -26,6 +26,7 @@ SIMPLE_CASES = [
("bbb", "aaa\nbbb\nccc", 1, 1 - 6 / 3),
("a", "a\nbb\nccc", 1, 1 - 5 / 1),
("bb", "a\nbb\nccc", 1, 1 - 4 / 2),
("abcd", "ab\ne", 1, 1 - 3 / 4),
]
COMPLEX_CASES = [
@ -135,7 +136,6 @@ def test_flexible_character_accuracy_xml(gt, ocr, first_line_score, all_line_sco
assert score == pytest.approx(all_line_score)
@pytest.mark.xfail(reason="Need to adapt performance details.")
@pytest.mark.parametrize(
"config,ocr",
[
@ -273,6 +273,8 @@ def test_remove_or_split(original, match, expected_lines):
[
*SIMPLE_EDITS,
(Part(text="a"), Part(text="b"), Distance(delete=1)),
(Part(text="ab"), Part(text="c"), Distance(delete=2)),
(Part(text="abc"), Part(text="d"), Distance(delete=3)),
(Part(text="aaa"), Part(text="bbb"), Distance(delete=3)),
(Part(text="aaabbbaaa"), Part(text="bbb"), Distance(match=3)),
(Part(text="bbb"), Part(text="aaabbbaaa"), Distance(match=3)),