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

fix: improve PP-StructureV3 structure preservation for complex diagrams

Browse Source 
- Fix parsing_res_list field mapping (block_label, block_content, block_bbox)
- Add fine-grained PP-StructureV3 configuration parameters
- Lower detection thresholds (0.5→0.2) for more sensitive element detection
- Use 'small' merge mode instead of default to minimize bbox merging
- Add layout_nms, unclip_ratio, text_det thresholds for better control
- Result: Doubled element detection from 6 to 12 elements on complex diagrams

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

Co-Authored-By: Claude <noreply@anthropic.com>
This commit is contained in:
egg
2025-11-25 08:53:37 +08:00
parent 4325d024a7
commit a659e7ae00
3 changed files with 73 additions and 17 deletions
Download Patch File Download Diff File Expand all files Collapse all files

66
backend/app/services/pp_structure_enhanced.py
View File

@@ -34,6 +34,7 @@ class PPStructureEnhanced:
# Mapping from PP-StructureV3 types to our ElementType
ELEMENT_TYPE_MAPPING = {
'title': ElementType.TITLE,
'paragraph_title': ElementType.TITLE, # PP-StructureV3 block_label
'text': ElementType.TEXT,
'paragraph': ElementType.PARAGRAPH,
'figure': ElementType.FIGURE,
@@ -107,24 +108,45 @@ class PPStructureEnhanced:
# Try to access parsing_res_list (the complete structure)
parsing_res_list = None
# Method 1: Direct access to json attribute
# Method 1: Direct access to json attribute (check both top-level and res)
if hasattr(page_result, 'json'):
result_json = page_result.json
if isinstance(result_json, dict) and 'parsing_res_list' in result_json:
parsing_res_list = result_json['parsing_res_list']
logger.info(f"Found parsing_res_list with {len(parsing_res_list)} elements")
if isinstance(result_json, dict):
# Check top-level
if 'parsing_res_list' in result_json:
parsing_res_list = result_json['parsing_res_list']
logger.info(f"Found parsing_res_list at top level with {len(parsing_res_list)} elements")
# Check inside 'res' (new structure in paddlex)
elif 'res' in result_json and isinstance(result_json['res'], dict):
if 'parsing_res_list' in result_json['res']:
parsing_res_list = result_json['res']['parsing_res_list']
logger.info(f"Found parsing_res_list inside 'res' with {len(parsing_res_list)} elements")
# Method 2: Try to access as attribute
# Method 2: Try direct dict access (LayoutParsingResultV2 inherits from dict)
elif isinstance(page_result, dict):
if 'parsing_res_list' in page_result:
parsing_res_list = page_result['parsing_res_list']
logger.info(f"Found parsing_res_list via dict access with {len(parsing_res_list)} elements")
elif 'res' in page_result and isinstance(page_result['res'], dict):
if 'parsing_res_list' in page_result['res']:
parsing_res_list = page_result['res']['parsing_res_list']
logger.info(f"Found parsing_res_list inside page_result['res'] with {len(parsing_res_list)} elements")
# Method 3: Try to access as attribute
elif hasattr(page_result, 'parsing_res_list'):
parsing_res_list = page_result.parsing_res_list
logger.info(f"Found parsing_res_list attribute with {len(parsing_res_list)} elements")
# Method 3: Check if result has to_dict method
# Method 4: Check if result has to_dict method
elif hasattr(page_result, 'to_dict'):
result_dict = page_result.to_dict()
if 'parsing_res_list' in result_dict:
parsing_res_list = result_dict['parsing_res_list']
logger.info(f"Found parsing_res_list in to_dict with {len(parsing_res_list)} elements")
elif 'res' in result_dict and isinstance(result_dict['res'], dict):
if 'parsing_res_list' in result_dict['res']:
parsing_res_list = result_dict['res']['parsing_res_list']
logger.info(f"Found parsing_res_list in to_dict['res'] with {len(parsing_res_list)} elements")
# Process parsing_res_list if found
if parsing_res_list:
@@ -207,27 +229,41 @@ class PPStructureEnhanced:
elements = []
for idx, item in enumerate(parsing_res_list):
# Extract element type
element_type = item.get('type', 'text').lower()
# Debug: log the structure of the first item
if idx == 0:
logger.info(f"First parsing_res_list item structure: {list(item.keys()) if isinstance(item, dict) else type(item)}")
logger.info(f"First parsing_res_list item sample: {str(item)[:500]}")
# Extract element type (check both 'type' and 'block_label')
element_type = item.get('type', '') or item.get('block_label', 'text')
element_type = element_type.lower()
mapped_type = self.ELEMENT_TYPE_MAPPING.get(
element_type, ElementType.TEXT
)
# Extract bbox (layout_bbox has the precise coordinates)
layout_bbox = item.get('layout_bbox', [])
if not layout_bbox and 'bbox' in item:
layout_bbox = item['bbox']
# Extract bbox (check multiple possible keys)
layout_bbox = (
item.get('layout_bbox', []) or
item.get('block_bbox', []) or
item.get('bbox', [])
)
# Ensure bbox has 4 values
if len(layout_bbox) >= 4:
bbox = layout_bbox[:4] # [x1, y1, x2, y2]
else:
bbox = [0, 0, 0, 0] # Default if bbox missing
logger.warning(f"Element {idx} has invalid bbox: {layout_bbox}")
# Extract content
content = item.get('content', '')
# Extract content (check multiple possible keys)
content = (
item.get('content', '') or
item.get('block_content', '') or
''
)
# Additional fallback for content in 'res' field
if not content and 'res' in item:
# Some elements have content in 'res' field
res = item.get('res', {})
if isinstance(res, dict):
content = res.get('content', '') or res.get('text', '')