TableExtractor

A minimal, research-focused pipeline for extracting tables from PDFs using state-of-the-art DETR-based models (Table Transformer) and simple postprocessing. This project is intended as a quick testbed and reference for table extraction, not as a production-ready solution.

---

Want to play around with the model? You can use our Google Colab notebook to experiment interactively. The notebook includes a nice explanation and is easy to copy and use in your own Google Drive.

🚀 Quick Start

1. Run with Docker (Recommended)

Requirements:

• Docker installed

# Build and run the API (serves on http://localhost:5000)
docker-compose up --build

• PDF files placed in pdfs/ will be accessible inside the container.
• Extracted images will be saved to images/.

2. Run with Python Virtual Environment

Requirements:

• Python 3.10+
• pip

# Create and activate a virtual environment
python3 -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Install dependencies
pip install --upgrade pip
pip install -r requirements.txt

# Download models (optional, will auto-download on first run)
python local_model.py

# Start the API
python src/app/main.py

---

🖥️ Frontend

A simple static HTML frontend is provided in frontend/index.html.
Open it in your browser and connect to the backend at http://localhost:5000.

---

🧠 How It Works

Main Pipeline

1. PDF Upload:
   User uploads a PDF via the frontend or API.

2. PDF to Images:
   Each page is converted to an image using PyMuPDF (src/app/pdf_reader.py).

3. Table Detection & Structure Recognition:

   • TD Model: Table Transformer Detection
     Detects tables and rotated tables in the image.
   • TSR Model: Table Structure Recognition
     Recognizes table structure (rows, columns, headers, spanning cells).

4. OCR:
   Each detected cell is read using EasyOCR. We suggest switching to another more porewfull OCR model.

5. Postprocessing:

   • Detected bounding boxes are mapped to a grid.
   • OCR results are assembled into a 2D array (list of lists of strings).

6. API Output:
   Returns the extracted table(s) as JSON.

---

🏷️ Model Output Classes

Table Detection (TD) Model

• "table": Standard table
• "table rotated": Rotated table
• "no object": No table detected

Table Structure Recognition (TSR) Model

• 0: 'table'
• 1: 'table column'
• 2: 'table row'
• 3: 'table column header'
• 4: 'table projected row header'
• 5: 'table spanning cell'

---

📦 API Endpoints

• POST /process
  Upload a PDF file (pdf field).
  Returns:

  {
    "data": [
      [
        ["cell1", "cell2", ...],
        ...
      ]
    ]
  }

• GET /health
  Health check.

---

🗂️ Data Postprocessing Logic

• Bounding boxes for rows, columns, and spanning cells are detected.
• A grid is constructed by intersecting row and column boxes.
• Spanning cells are mapped to all grid positions they cover.
• Each cell is cropped and OCR is applied.
• The result is a rectangular 2D array, padded as needed.

Note: For robust postprocessing, thresholding, and extracting meaning from model predictions, we strongly recommend referring to the official Microsoft Table Transformer postprocess.py. Their code covers many edge cases and implements a much more comprehensive logic for table structure extraction. As this was not the main scope of our project, we suggest using their approach for production or research-grade extraction. They also provide clear instructions for training and fine-tuning the model.

---

⚠️ Limitations & Recommendations

• This repo is a quick testbed, not a production system.
• For complex tables, multi-page tables, or robust extraction, refer to the above projects.
• Postprocessing is intentionally simple and may fail on edge cases.
• We suggest to use better OCR model, as easyOCR can fail in lots of casese.

---

🖼️ Example: Complex Tables

For examples of complex table extraction, see the RAGFlow project and its documentation.

---

📁 Project Structure

TableExtractorDETR/
  ├── src/app/           # Main backend code (Flask API, inference, PDF reader)
  ├── models/            # Downloaded model weights (auto-populated)
  ├── pdfs/              # Input PDFs
  ├── images/            # Output images from PDF pages
  ├── frontend/          # Static HTML frontend
  ├── requirements.txt   # Python dependencies
  ├── Dockerfile         # Docker build
  ├── docker-compose.yml # Docker Compose config
  └── local_model.py     # Script to pre-download models

---

📚 References

• Table Transformer (TATR) Paper
• Microsoft Table Transformer GitHub
• RAGFlow

---

License

MIT

---