PEC-Hackathon

Document Processing API -- README

This repository contains a FastAPI-based backend for document processing using YOLO-based layout detection, OCR-based table extraction, and OneChart-based chart analysis.

---

🚀 Features

1. Document Layout Detection (YOLOv8)

• Detects tables, charts, images, and figures from PDF or image files.
• Crops and classifies regions using a custom YOLO model.
• Preloads the model at startup for fast inference.

2. Table Extraction with OCR (EasyOCR)

• Applies OCR on detected table regions.
• Auto-detects table rows/columns using bounding box clustering.
• Returns structured JSON output ready for UI display.

3. Chart Analysis Using OneChart (VLM)

• Extracts key information from charts using the OneChart transformer model.
• Gracefully handles CUDA errors and switches to CPU fallback.
• Uses timeouts to prevent long-running GPU operations.

⚠️ Note:
ChartOCR / OneChart currently fails or times out due to insufficient GPU VRAM.
On systems with low compute resources, chart analysis will return "status": "failed" or "timeout".

---

📦 Installation

1. Clone the repository

git clone <repo-url>
cd <project-folder>

2. Create and activate environment

You must install dependencies manually or via requirements.txt.

Example:

conda create -n doc-process python=3.10
conda activate doc-process
pip install -r requirements.txt

3. Install Poppler (for PDF processing)

Required for pdf2image.

Ubuntu:

sudo apt install poppler-utils

Windows:

Download from:
https://github.com/oschwartz10612/poppler-windows

---

▶️ Running the API

uvicorn main:app --reload

Server runs at:

http://127.0.0.1:8000

---

📤 API Endpoints

1. Health Check

GET /health

2. Upload & Process a File

POST /upload_and_process/

Payload: - PDF, PNG, JPG files - Optional flag analyze_charts=true

Returns: - Layout detections\

• Extracted table data\
• Chart images & analysis (if enabled)\
• Summary statistics

---

📁 Project Structure

main.py                 ← Main FastAPI application  
models/                 ← YOLO model files (custom)  
utils/                  ← OCR + processing utilities  

---

⚠️ Known Issues

1. ❗ OneChart fails on GPU with limited VRAM

• Causes CUDA OOM or inference timeout
• CPU fallback works but is very slow
• Expected output: "status": "failed", "timeout", or "skipped"

2. ❗ EasyOCR is slow on large tables

• Consider switching to PaddleOCR or Tesseract if speed is required.

---

🔧 Customization

Change YOLO model path:

Set environment variable:

export YOLO_MODEL_PATH="path/to/model.pt"

Disable chart analysis:

POST /upload_and_process/?analyze_charts=false

---

📄 License

MIT License (modify as per your project needs)

---

If you need a documentation site, Postman collection, or frontend integration, I can generate those as well.