ANPR — Automatic Number Plate Recognition

A full-stack AI application that detects and reads vehicle number plates from images. Upload a photo through the web interface and get the plate text back in seconds.

---

How It Works

Image Upload → YOLOv8 plate detection → EasyOCR text extraction → Result + DB storage

1. YOLOv8 (best.pt) detects the license plate and returns a tight bounding box
2. EasyOCR reads the text from the cropped plate region using multi-variant preprocessing
3. Positional correction fixes common OCR misreads based on Indian plate format (SS DD LL DDDD)
4. Results are saved to SQLite and returned to the frontend

---

Tech Stack

Layer	Technology
Backend	Python, FastAPI, Uvicorn
AI / CV	YOLOv8 (Ultralytics), EasyOCR, OpenCV
Database	SQLite, SQLAlchemy
Frontend	Next.js 16, React 19, Tailwind CSS v4, TypeScript

---

Project Structure

ANPR/
├── backend/
│   ├── api.py              # FastAPI routes (/scan-plate, /history)
│   ├── core/
│   │   ├── detector.py     # YOLOv8 plate detection
│   │   ├── ocr.py          # EasyOCR + preprocessing + correction
│   │   └── pipeline.py     # Orchestrates detection → OCR
│   ├── database.py         # SQLAlchemy engine + session
│   ├── models.py           # ScanResult DB model
│   └── requirements.txt
├── frontend/
│   └── src/app/
│       ├── page.tsx        # Main UI (upload, results, history)
│       └── layout.tsx
├── images/                 # Sample test images
├── outputs/                # CLI output images with bounding boxes
├── uploads/                # Saved uploaded images
├── main.py                 # CLI test script
├── best.pt                 # Fine-tuned YOLOv8 plate detection model
├── yolov8n.pt              # Base YOLOv8n weights (used for retraining)
├── haarcascade_russian_plate_number.xml
└── anpr_data.db            # SQLite database

---

Getting Started

Prerequisites

• Python 3.10+
• Node.js 18+
• Tesseract OCR (Windows): Download installer — install to C:\Program Files\Tesseract-OCR\

1. Backend

# Create and activate virtual environment (recommended)
python -m venv venv
venv\Scripts\activate        # Windows
source venv/bin/activate     # macOS/Linux

# Install dependencies
pip install -r backend/requirements.txt

# Start the API server
uvicorn backend.api:app --reload

API runs at http://127.0.0.1:8000
Swagger docs at http://127.0.0.1:8000/docs

2. Frontend

cd frontend
npm install
npm run dev

Open http://localhost:3000

3. CLI (no server needed)

python main.py

Processes images/car.jpg and saves the annotated result to outputs/predicted_result.jpg.

---

API Endpoints

Method	Endpoint	Description
POST	/api/v1/scan-plate	Upload an image, returns detected plate text
GET	/api/v1/history	Returns last 20 scan records
GET	/	Health check

Example response from /api/v1/scan-plate

{
  "success": true,
  "processing_time_ms": 4200.5,
  "plates_detected": 1,
  "results": [
    {
      "text": "HR26DK8337",
      "confidence": 0.87,
      "warning": null,
      "box": { "x": 120, "y": 310, "w": 390, "h": 100 }
    }
  ]
}

---

Known Limitations

• Tilted or angled plates — perspective distortion reduces OCR accuracy. Plates at more than ~15° angle may produce incorrect characters. Perspective correction (deskewing) is not yet implemented.
• Partially visible plates — if characters are obscured, out of frame, or covered, the result will be incomplete. The API returns a warning when fewer than 6 characters are detected.
• Low confidence reads — results below 50% confidence are flagged with a warning in the UI. These should be treated as unreliable.
• Indian plate format assumed — the positional character correction (letter/digit mapping) is tuned for the Indian SS DD LL DDDD format. International plates may get incorrectly corrected.
• CPU speed — EasyOCR runs 3 preprocessing variants per plate on CPU, taking 15–30 seconds per image. A CUDA GPU reduces this to under 3 seconds.
• best.pt model — trained on ~1100 Indian plate images. Accuracy may vary on plates from other regions, heavily stylized plates, or very low resolution images.

---

Retraining the Plate Detection Model

The best.pt model was trained on Google Colab using a Roboflow dataset. To retrain or improve it:

1. Get a labeled plate dataset from Roboflow Universe in YOLOv8 format
2. Run on Google Colab (free T4 GPU):

from ultralytics import YOLO
model = YOLO("yolov8n.pt")
model.train(data="data.yaml", epochs=50, imgsz=640, batch=16, device=0)

3. Download runs/detect/train/weights/best.pt and replace the existing best.pt in the project root

---

Environment Variables

Variable	Default	Description
TESSERACT_CMD	C:\Program Files\Tesseract-OCR\tesseract.exe	Override Tesseract path
NEXT_PUBLIC_API_URL	http://127.0.0.1:8000	Backend URL for the frontend

Set NEXT_PUBLIC_API_URL in frontend/.env.local if your backend runs on a different port.

---

License

This project is open-source.