PDF Structured Outline Extractor (Adobe Hackathon 1A)

Overview

This solution automatically extracts a hierarchical outline (Title, H1, H2, H3 headings and their page numbers) from PDF documents, outputting results as specification-compliant JSON files.
Optimized for accuracy, speed, and generality—no PDF-specific rules or network calls.

Features

• Accurate heading detection using a multi-feature approach (font size, boldness, position, font name, sentence structure).
• Handles standard and complex PDFs; robust to line fragmentation, noisy layout, and various heading conventions.
• Conforms to hackathon requirements
  • Max runtime: ≤10 seconds (50p PDF)
  • Model/code size: <200MB
  • CPU only, offline, runs in Docker (amd64)
  • Processes all /app/input/*.pdf, outputs /app/output/*.json
• Bonus: Multilingual normalization for improved handling of non-English PDFs.

Project Structure

.
├── src/
│   └── extract_outline.py        # Main extraction script
├── requirements.txt              # Python dependencies (PyMuPDF)
├── Dockerfile                    # For reproducible builds/execution
├── input/                        # Place your PDF files here (mounted in docker)
├── output/                       # JSON output appears here (mounted in docker)
└── README.md                     # This file

Installation & Usage

1. Build the Docker image (Required)

Ensure Docker is installed and running.
From the project root:

docker build --platform linux/amd64 -t pdf_outline_extractor_1a .

2. Prepare Input/Output Folders

Place your PDF files (≤50 pages each) in the input/ directory:

mkdir -p input output
# Copy or move PDF files to input/

3. Run the Extractor

docker run --rm \
  -v "$PWD/input:/app/input" \
  -v "$PWD/output:/app/output" \
  --network none \
  pdf_outline_extractor_1a

• Each filename.pdf in input/ produces a corresponding filename.json in output/.

Output Format Example

{
  "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

• PDF Parsing: Uses PyMuPDF for fast extraction of text and layout metadata, fully offline.
• Heading Detection:
  • Combines font size (statistical z-scoring), bold/italic style, font name, text length, and punctuation cues for robust section detection.
  • Adjacent heading fragments (caused by PDF line breaks) are merged using y-position and font similarity.
• Title Extraction:
  • Selects the largest, top-most text block on the first page, excluding body lines and fragments.
• Outline Structuring:
  • Each detected heading is assigned a hierarchical level (H1, H2, H3) and deduplicated for clarity.
• Multilingual Handling:
  • Unicode normalization and whitespace cleanup included; basic support for non-English character sets.

Dependencies

• Python 3.8+
• PyMuPDF==1.23.0
• Standard libraries: os, json, re, unicodedata

To install manually:

pip install -r requirements.txt

Constraints & Best Practices

• Handles any PDF (up to 50 pages); no per-document hardcoding.
• Model/code footprint <200MB.
• Strictly CPU-only, no GPU or network dependency.
• All processing occurs inside the container.
• For best accuracy, test on a variety of document types.

Troubleshooting

• If output JSON is empty, ensure your PDF uses selectable, extractable text (not only scanned images).
• For edge cases (e.g., very fragmented headings), tune merging thresholds in the script as needed.

Contact

For queries or feedback, please contact Yuvraj Kumar yuvraj.kr2022@gmail.com or open an issue in your private workspace.

Built for Adobe India Hackathon 2025 — Connecting the Dots, Round 1A.
Ready for direct evaluation via the provided Docker instructions.

---