ADOBE ROUND 1A

A lightweight PDF outline extractor that identifies document headings and generates a JSON outline of title, levels, text, and page number.

• Multilingual Support: Works seamlessly on PDFs in multiple languages (e.g., English, French, Chinese, German).
• High Performance: Processes documents up to 50 pages in under 5–6 seconds on average.

Approach

1. Text and Feature Extraction

   • Use PyMuPDF (fitz) to parse each PDF page into text blocks and lines.
   • For each line, extract raw features such as font size, position (x0, y0), text length, uppercase ratio, bold/italic flags, numeric or punctuation patterns, and title-case flag.

2. Batch Predictions for Heading Detection

   • Aggregate all line features into a single Pandas DataFrame.
   • Perform a single heading_model.predict() over the entire DataFrame to flag heading lines.
   • This reduces per-line DataFrame construction and model-call overhead, improving performance on large documents.

3. Merge Split/Overlapping Lines

   • Iteratively merge consecutive fragments if they have the same font size and their vertical positions (y0) fall within a threshold.
   • Prevent duplicate text fragments by checking if the new fragment is already contained in the current heading entry.

Post-Processing

4. Map Font Sizes to Heading Levels (H1–H6)

   • Identify the unique font sizes among detected headings in descending order.
   • Assign the largest font size to H1, the second-largest to H2, and so on up to H6.

5. De-duplicate and Finalize Outline

   • Remove any remaining duplicate heading texts.
   • Select the first H1 as the document title.
   • Produce a JSON outline containing a title string and an array of { level, text, page } objects.

Models & Libraries Used

• Models

  • heading_classifier.pkl: Binary classifier (e.g., Random Forest or Logistic Regression) to detect heading vs. body text.
  • level_classifier.pkl: Multiclass model for heading levels; currently provided for extensibility but not used in the main batch pipeline.

• Python Libraries

  • PyMuPDF (fitz) for PDF parsing.
  • Pandas for DataFrame-based batching.
  • scikit-learn for loading and predicting with pre-trained models.
  • joblib for efficient model serialization and loading.

Installation

1. Clone the Repository

git clone https://github.com/AdamyaSingh7/AdobeChallenge_1A.git
cd AdobeChallenge_1a

2. Set Up Prerequisites

• Ensure Docker is installed (version 20+ recommended).
• (Optional) If running locally, use Python 3.8+ and install dependencies:

pip install -r requirements.txt

Build & Run

Build Docker Image

In Bash (build):

docker build --platform linux/amd64 -t adobe1a-outline-extractor:latest .

In PowerShell (build):

docker build --platform linux/amd64 -t adobe1a-outline-extractor:latest .

⚠️ Note: The build process takes approximately 1 to 1.5 minutes, depending on your system.

Run Container

In Bash (run):

docker run --rm \
  -v $(pwd)/input:/app/input \
  -v $(pwd)/output:/app/output \
  --network none \
  adobe1a-outline-extractor:latest

In PowerShell (run):

docker run --rm `
  -v "${PWD}\input:/app/input" `
  -v "${PWD}\output:/app/output" `
  --network none `
  adobe1a-outline-extractor:latest

The tool will process all PDFs in /app/input and write JSON outlines to /app/output.

Note: The /input folder also includes extra test PDFs beyond the Adobe-provided samples:

• A 50-page document to benchmark performance.
• A German-language PDF to validate multilingual support.