egg/OCR
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix: add IoU text coverage check and page boundary validation

Browse Source 
Vector rectangles:
- Add page boundary check (skip out-of-bounds rectangles)
- Clip rectangles to page boundaries

Covering images:
- Add page boundary check (skip out-of-bounds images)
- Add IoU-based text coverage verification
- Only report images that actually cover text (>= 50% word coverage)
- Add covered_text_count to detection results

This reduces false positives from black logos or decorative images
that don't actually cover any text content.

Test results (edit3.pdf):
- Before: 10 covering images detected
- After: 6 covering images detected (4 filtered - no text coverage)

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>
This commit is contained in:
egg
2025-12-04 07:48:38 +08:00
parent 1c3c37bce0
commit 9437387ef1
1 changed files with 54 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files

57
backend/app/services/direct_extraction_engine.py
View File

@@ -1941,6 +1941,7 @@ class DirectExtractionEngine:
List of dicts with covered text info: {'text', 'bbox', 'coverage', 'color_type'}
"""
covered_words = []
page_rect = page.rect # Page boundaries
# Get all drawings and find solid-filled rectangles
drawings = page.get_drawings()
@@ -1960,6 +1961,13 @@ class DirectExtractionEngine:
if fitz_rect.width < 5 or fitz_rect.height < 5:
continue
# Skip rectangles completely outside page boundaries
if not fitz_rect.intersects(page_rect):
continue
# Clip rectangle to page boundaries
fitz_rect = fitz_rect & page_rect
# Detect white rectangles (white-out / correction tape)
# Must be pure white (>= 0.98) to avoid false positives from light backgrounds
if r >= 0.98 and g >= 0.98 and b >= 0.98:
@@ -1975,7 +1983,7 @@ class DirectExtractionEngine:
# Log detected covering rectangles by type
white_count = sum(1 for _, t in covering_rects if t == 'white')
black_count = sum(1 for _, t in covering_rects if t == 'black')
logger.debug(f"Page {page_num}: Found {len(covering_rects)} covering rectangles "
logger.debug(f"Page {page_num}: Found {len(covering_rects)} potential covering rectangles "
f"(white: {white_count}, black/redaction: {black_count})")
# Get all text words with bounding boxes
@@ -2013,7 +2021,12 @@ class DirectExtractionEngine:
def _detect_covering_images(self, page: fitz.Page, doc: fitz.Document, page_num: int) -> List[Dict]:
"""
Detect embedded images that are mostly black/white (likely covering/redaction).
Detect embedded images that are mostly black/white AND actually cover text.
Only reports images that:
1. Are mostly solid black or white
2. Are within page boundaries
3. Actually overlap with text content (IoU check)
Args:
page: PyMuPDF page object
@@ -2021,9 +2034,10 @@ class DirectExtractionEngine:
page_num: Page number for logging
Returns:
List of dicts with covering image info: {'bbox', 'color_type', 'avg_color'}
List of dicts with covering image info: {'bbox', 'color_type', 'avg_color', 'covered_text_count'}
"""
covering_images = []
page_rect = page.rect # Page boundaries
try:
# Get all images on the page with their positions
@@ -2032,6 +2046,9 @@ class DirectExtractionEngine:
if not image_list:
return covering_images
# Get all text words for coverage check
words = page.get_text("words") # (x0, y0, x1, y1, word, block_no, line_no, word_no)
for img_info in image_list:
xref = img_info[0]
width = img_info[2]
@@ -2076,13 +2093,38 @@ class DirectExtractionEngine:
if color_type:
# Get image position on page
# We need to find the image rectangle on the page
for img_rect in page.get_image_rects(xref):
# Skip images completely outside page boundaries
if not img_rect.intersects(page_rect):
continue
# Clip image rect to page boundaries
clipped_rect = img_rect & page_rect
# Check if image actually covers any text (IoU check)
covered_text_count = 0
for word_info in words:
word_rect = fitz.Rect(word_info[:4])
word_area = word_rect.width * word_rect.height
if word_area <= 0:
continue
intersection = word_rect & clipped_rect
if not intersection.is_empty:
intersection_area = intersection.width * intersection.height
coverage_ratio = intersection_area / word_area
# Count as covered if >= 50% of word is under the image
if coverage_ratio >= 0.5:
covered_text_count += 1
# Only report if image actually covers text
if covered_text_count > 0:
covering_images.append({
'bbox': tuple(img_rect),
'bbox': tuple(clipped_rect),
'color_type': color_type,
'avg_color': (avg_r, avg_g, avg_b),
'size': (width, height)
'size': (width, height),
'covered_text_count': covered_text_count
})
except Exception as e:
@@ -2092,8 +2134,9 @@ class DirectExtractionEngine:
if covering_images:
black_count = sum(1 for c in covering_images if c['color_type'] == 'image_black')
white_count = sum(1 for c in covering_images if c['color_type'] == 'image_white')
total_covered = sum(c.get('covered_text_count', 0) for c in covering_images)
logger.debug(f"Page {page_num}: Found {len(covering_images)} covering images "
f"(black: {black_count}, white: {white_count})")
f"(black: {black_count}, white: {white_count}, covering {total_covered} text regions)")
except Exception as e:
logger.warning(f"Page {page_num}: Failed to detect covering images: {e}")