Adobe India Hackathon 2025 – Round 1A: Structured Outline Extraction

Problem Statement

Challenge 1A required us to extract a structured document outline from unstructured PDF files. This includes:

• Extracting the document title from the largest font block.
• Identifying section headings using a combination of font features and layout rules.
• Categorizing headings into hierarchical levels (H1, H2, H3) using font size or numbered patterns.
• Saving the extracted data in a clean, structured JSON format.

---

irectory Structure

├── app/
│   ├── main.py                   # Main script to extract title + outline from all PDFs
│   ├── requirements.txt          # Dependencies (PyMuPDF, etc.)
│   ├── Dockerfile                # Docker setup for Round 1A
│   ├── input/                    # Place PDF files here
│   └── output/                   # JSON output files generated

---

Key Features

• Extracts clean document titles (even if spread across multiple lines).
• Detects headings based on:
  • Font size
  • Bold/Italic style
  • Numbered pattern (e.g., 1.2, 2.1.3)
• Ensures no duplicate or footer content is included.
• Tags headings with hierarchical level: H1, H2, H3.

---

Dependencies

Install required libraries:

pip install -r requirements.txt

Libraries used:

• PyMuPDF (fitz) – for reading PDFs and fonts
• re, os, json – for processing and formatting

---

🛠️ How to Run Locally

🔧 Step 1: Place PDFs

Put your PDF files inside the input/ folder.

Step 2: Run the script

python main.py

This will create a .json output for each PDF inside the output/ folder.

---

Run with Docker

Step 1: Build the Docker image

docker build -t round1a-extractor .

Step 2: Run the container

docker run --rm -v ${PWD}/input:/app/input -v ${PWD}/output:/app/output round1a-extractor

Your output will be saved in the output/ folder.

---

Download / Submission Package

You can download this submission as a zip:

Download Round 1A Submission

---

---

Author

Strugglers

Feel free to reach out for any doubts or improvements!

---

Notes

• Round 1A is offline-compatible.
• No AI models are used for classification; rules are applied based on font metrics.
• Docker makes the setup universal across systems.

---