Skip to content

README.md

Latest commit

 

History

History
114 lines (88 loc) · 3.32 KB

README.md

File metadata and controls

114 lines (88 loc) · 3.32 KB

US Driver License Data Extraction

A multi-approach machine learning project for extracting text data (name, DOB, address, state) from US driver's license images using object detection and OCR techniques.

Overview

This repository explores multiple approaches for driver's license data extraction:

  • YOLO-v8 + EasyOCR: Production-ready pipeline combining YOLOv8 for field detection with EasyOCR for text recognition
  • YOLO-v5: Custom-trained YOLOv5 model for license field detection
  • Donut (Document Understanding Transformer): End-to-end transformer-based document parsing without OCR
  • pyTesseract: Traditional OCR approach using Tesseract engine
  • YOLO + TensorFlow 1.0: Legacy implementation using Darkflow

Repository Structure

.
├── DATASET/                    # Training datasets (US & India driver licenses)
├── Donut/
│   ├── CORD/                   # Fine-tuning notebooks and data preparation scripts
│   └── deployment/             # Gradio web app for Donut model inference
├── YOLO-v5/
│   ├── LegacyTrain/            # Training artifacts
│   └── yolov5/                 # YOLOv5 implementation
├── YOLO-v8/
│   ├── main.py                 # FastAPI REST API endpoint
│   ├── streamlit-app.py        # Streamlit web interface
│   ├── parseq/                 # Scene text recognition model
│   └── new_models/             # Trained model weights
├── pyTesseract_OCR/            # Tesseract-based extraction
└── YOLO___TF1.0/               # Legacy TensorFlow 1.x implementation

Quick Start

YOLOv8 + EasyOCR (Recommended)

  1. Install dependencies:

    cd YOLO-v8
pip install -r requirements.txt

  2. Run the Streamlit app:

    streamlit run streamlit-app.py

  3. Or start the FastAPI server:

    uvicorn main:app --reload

    API endpoint: POST /predict with image file

Donut Transformer

  1. Install dependencies:

    cd Donut/deployment
pip install -r requirements.txt

  2. Run the Gradio app:

    python gradio-app.py

YOLOv5

  1. Create conda environment:

    cd YOLO-v5
conda env create -f my_conda_env_yolov5.yml
conda activate YOLO-v5

  2. Training:

    python train.py --img 640 --batch 4 --epochs 100 --data train/US_DL.yaml --cfg train/custom_yolov5l.yaml --weights train_L/yolov5l.pt

  3. Inference:

    python detect.py --weights runs/train/{exp}/weights/best.pt --img 640 --conf 0.2 --source test/images

Tech Stack

  • Deep Learning: PyTorch, Ultralytics YOLOv8/v5, Hugging Face Transformers
  • OCR: EasyOCR, pyTesseract, PARSeq
  • Web Frameworks: FastAPI, Streamlit, Gradio
  • Computer Vision: OpenCV, PIL/Pillow
  • Other: PyTorch Lightning, CUDA support for GPU acceleration

Models

Approach Detection Text Recognition Use Case
YOLOv8 + EasyOCR YOLOv8 EasyOCR Production API/UI
Donut Transformer End-to-end Zero-OCR pipeline
YOLOv5 YOLOv5 - Field detection only
pyTesseract - Tesseract Basic OCR

Extracted Fields

  • Name
  • Date of Birth (DOB)
  • Address
  • State

License

See individual component licenses in YOLO___TF1.0/licenses/.