Benjamin_Boenisch/breakpilot-lehrer
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

refactor(ocr-pipeline): use left-edge alignment approach for sub-column detection

Browse Source 
Replace gap-based splitting with alignment-bin approach: cluster word
left-edges within 8px tolerance, find the leftmost bin with >= 10% of
words as the true column start, split off any words to its left as a
sub-column. This correctly handles both page references ("p.59") and
misread exclamation marks ("!" → "I") even when the pixel gap is small.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This commit is contained in:
Benjamin Admin
2026-03-02 18:56:38 +01:00
parent f13116345b
commit 7252f9a956
2 changed files with 87 additions and 97 deletions
Download Patch File Download Diff File Expand all files Collapse all files

52
klausur-service/backend/tests/test_cv_vocab_pipeline.py
View File

@@ -1177,7 +1177,7 @@ class TestRegionContentCheck:
# =============================================
class TestSubColumnDetection:
"""Tests for _detect_sub_columns() left-edge clustering."""
"""Tests for _detect_sub_columns() left-edge alignment detection."""
def _make_word(self, left: int, text: str = "word", conf: int = 90) -> dict:
return {'left': left, 'top': 100, 'width': 50, 'height': 20,
@@ -1191,27 +1191,46 @@ class TestSubColumnDetection:
)
def test_sub_column_split_page_refs(self):
"""Column with 3 'p.XX' left + 20 EN words right → split into 2."""
"""3 page-refs left + 30 vocab words right → split into 2.
The leftmost bin with >= 10% of words (i.e. >= 4) is the vocab bin
at left=250, so the 3 page-refs are outliers.
"""
content_w = 1000
# 3 page-ref words at left=100, 20 vocab words at left=250
page_words = [self._make_word(100, f"p.{59+i}") for i in range(3)]
vocab_words = [self._make_word(250, f"word{i}") for i in range(20)]
vocab_words = [self._make_word(250, f"word{i}") for i in range(30)]
all_words = page_words + vocab_words
geo = self._make_geo(x=80, width=300, words=all_words, content_w=content_w)
result = _detect_sub_columns([geo], content_w)
assert len(result) == 2, f"Expected 2 columns, got {len(result)}"
# Left sub-column should be narrower with fewer words
left_col = result[0]
right_col = result[1]
assert left_col.x < right_col.x
assert left_col.word_count == 3
assert right_col.word_count == 20
# Indices should be 0, 1
assert right_col.word_count == 30
assert left_col.index == 0
assert right_col.index == 1
def test_sub_column_split_exclamation_marks(self):
"""5 '!' (misread as I/|) left + 80 example words → split into 2.
Mirrors the real-world case where red ! marks are OCR'd as I, |, B, 1
at a position slightly left of the example sentence start.
"""
content_w = 1500
bang_words = [self._make_word(950 + i, chr(ord('I')), conf=60) for i in range(5)]
example_words = [self._make_word(975 + (i * 3), f"word{i}") for i in range(80)]
all_words = bang_words + example_words
geo = self._make_geo(x=940, width=530, words=all_words, content_w=content_w)
result = _detect_sub_columns([geo], content_w)
assert len(result) == 2
assert result[0].word_count == 5
assert result[1].word_count == 80
def test_no_split_uniform_alignment(self):
"""All words aligned at same position → no change."""
content_w = 1000
@@ -1228,7 +1247,6 @@ class TestSubColumnDetection:
content_w = 1000
words = [self._make_word(50, "a")] * 3 + [self._make_word(120, "b")] * 10
geo = self._make_geo(x=40, width=140, words=words, content_w=content_w)
# width_ratio = 140/1000 = 0.14 < 0.15
result = _detect_sub_columns([geo], content_w)
@@ -1241,7 +1259,6 @@ class TestSubColumnDetection:
right_words = [self._make_word(300, f"b{i}") for i in range(12)]
all_words = left_words + right_words
geo = self._make_geo(x=80, width=400, words=all_words, content_w=content_w)
# 8/20 = 0.4 >= 0.35 → no split
result = _detect_sub_columns([geo], content_w)
@@ -1250,26 +1267,23 @@ class TestSubColumnDetection:
def test_sub_column_reindexing(self):
"""After split, indices are correctly 0, 1, 2 across all columns."""
content_w = 1000
# First column: no split
# First column: no split (all words at same alignment)
words1 = [self._make_word(50, f"de{i}") for i in range(10)]
geo1 = ColumnGeometry(index=0, x=30, y=50, width=200, height=500,
word_count=10, words=words1, width_ratio=0.2)
# Second column: will split
# Second column: will split (3 outliers + 30 main)
page_words = [self._make_word(400, f"p.{i}") for i in range(3)]
en_words = [self._make_word(550, f"en{i}") for i in range(15)]
en_words = [self._make_word(550, f"en{i}") for i in range(30)]
geo2 = ColumnGeometry(index=1, x=380, y=50, width=300, height=500,
word_count=18, words=page_words + en_words, width_ratio=0.3)
word_count=33, words=page_words + en_words, width_ratio=0.3)
result = _detect_sub_columns([geo1, geo2], content_w)
assert len(result) == 3
assert [g.index for g in result] == [0, 1, 2]
# First column unchanged
assert result[0].word_count == 10
# Sub-column (page refs)
assert result[1].word_count == 3
# Main column (EN words)
assert result[2].word_count == 15
assert result[2].word_count == 30
def test_no_split_too_few_words(self):
"""Column with fewer than 5 words → no split attempted."""
@@ -1283,10 +1297,10 @@ class TestSubColumnDetection:
assert len(result) == 1
def test_no_split_single_minority_word(self):
"""Only 1 word in minority cluster → no split (need >= 2)."""
"""Only 1 word left of column start → no split (need >= 2)."""
content_w = 1000
minority = [self._make_word(100, "p.59")]
majority = [self._make_word(300, f"w{i}") for i in range(20)]
majority = [self._make_word(300, f"w{i}") for i in range(30)]
geo = self._make_geo(x=80, width=350, words=minority + majority, content_w=content_w)
result = _detect_sub_columns([geo], content_w)