Extract a structured outline from any PDF document. This includes:
- Title
- Headings (H1, H2, H3)
- Page numbers
Output format: json { "title": "Understanding AI", "outline": [ { "level": "H1", "text": "Introduction", "page": 1 }, { "level": "H2", "text": "What is AI?", "page": 2 }, { "level": "H3", "text": "History of AI", "page": 3 } ] } 🧠 Approach We developed a robust pipeline that uses both visual and semantic cues to extract headings from PDFs. The process includes:
Text Extraction: Extract raw text and layout metadata using PyMuPDF (fitz).
Candidate Identification: Use heuristics (title casing, font size, numbering) and NER (spaCy) to propose heading candidates.
Matching: Fuzzy match candidates against extracted spans with font/style metadata.
Filtering: Remove repeated headers/footers using positional clustering.
Merging: Merge adjacent heading spans with same styling.
Scoring & Ranking: Compute importance score using:
Font size
Font weight
Box enclosure (e.g., if it's inside a bounding box)
Tagging: Top-scored spans are tagged as H1/H2/H3 accordingly.
Output: Structured JSON is generated with proper hierarchy.
🧱 Tech Stack Python 3.11
PyMuPDF
spaCy (en_core_web_sm)
🐳 Docker Instructions
🔨 Build
bash
Copy
Edit
docker build --platform linux/amd64 -t headingextractor:latest .
-v "$(pwd)/input:/app/input"
-v "$(pwd)/output:/app/output"
--network none
headingextractor:latest
The container will:
Read all PDFs from /app/input
Write JSON outlines to /app/output
✅ Constraints Satisfied Constraint Status ≤ 10s for 50-page PDF ✅ CPU only ✅ Model size ≤ 200MB ✅ Offline (no network) ✅ Output JSON format ✅
📁 Directory Structure css Copy Edit . ├── round1A.py ├── main.py ├── requirements.txt ├── Dockerfile ├── input/ └── output/ 📌 Notes Works for a wide range of document layouts.
No hardcoded logic or file-specific assumptions.