Automatic Newspaper Reading System

Critical data was locked in print

Despite the digital revolution, a significant portion of the world's information remains locked in physical formats. For media monitoring agencies, historical archives, and legal firms, the inability to rapidly search and analyze print newspapers represents a critical bottleneck. Manual digitization is prohibitively expensive and slow, while traditional OCR tools fail when confronted with the complex, multi-column layouts and mixed media - text, images, ads - typical of broadsheet journalism.

Newspapers are visually the hardest documents to parse

A single newspaper page might contain five different articles, three advertisements, and a crossword puzzle, all arranged in a non-linear grid. A standard left-to-right OCR scan produces a garbled stream of text where headlines merge with unrelated columns. The engineering challenge was twofold: accurately segment the page into discrete semantic units, then extract clean text from each segment despite varying fonts, low print quality, and background noise.

A hybrid deep learning pipeline

TendersLab engineered an end-to-end pipeline that treats newspaper reading as a multi-stage computer vision problem.

1. Semantic segmentation with Mask R-CNN

A custom Mask R-CNN model trained on annotated newspaper pages acts as the system's "eyes" - identifying bounding boxes for articles, headlines, images, and advertisements. Crucially, it understands the hierarchical relationship between a headline and its associated body text, even when they span multiple columns and contain mixed-direction text.

2. OCR and error correction

Each segmented text block is passed to a Tesseract-based OCR engine optimized for print typography. A post-processing NLP layer corrects common errors caused by newsprint bleed-through or folds - using language models to fix OCR typos (e.g., 'rn' confused with 'm') based on surrounding context.

3. Structured data extraction

The final stage converts raw text into structured JSON. For advertisements, the system extracts phone numbers, dates, prices, and named entities using Regex and Named Entity Recognition (NER) - turning a static classified ad image into a searchable, filterable database entry.

From paper archive to live database

• 95% reduction in processing time: A task that took a human analyst 20 minutes is completed by the pipeline in under 60 seconds.
• High-fidelity archiving: The system preserves the visual context of the original page while making all content fully searchable and indexable.
• 24/7 automated ingestion: New editions are processed automatically as they are scanned - no human-in-the-loop required for standard pages.