The Challenge
Overlapping articles, irregular column layouts, and advertisements made rule-based segmentation unreliable (50% accuracy). Manual annotation cost $2 per page.
Our Solution
Fine-tuned a Mask R-CNN model on 10K labeled newspaper pages, with rules for ad/classifieds filtering. Added a post-processing step to merge fragmented regions and infer reading order.
Technologies Used
Results & Impact
- Achieved 94% segmentation accuracy across 19th–20th century newspapers
- Reduced processing costs from $2/page to $0.05/page
- Improved downstream OCR accuracy by 22% via clean article isolation