Lightweight License Plate Character Recognition

This project implements a character recognition pipeline for license plates using PyTorch and OpenCV. It supports:

---

Features

• End-to-end character recognition pipeline
• CNN-based classifier
• Noise-resilient image preprocessing & augmentation
• Batch prediction and automatic annotation
• Tailored for Australian license plates

---

Example

Input Image 1:

Predicted Output 1:

Input Image 2:

Predicted Output 2:

---

Model Training

The model is trained on a subset of the Chars74K dataset (EnglishFnt.tgz).
Only characters visually similar to Victorian license plates were manually retained to ensure performance on real-world data.

Dataset Preparation

Download EnglishFnt.tgz

Extract to ./dataset/train/

Manually remove non-standard / stylized characters not visually aligned with Victorian plates

Project Structure

.
├── train.py          # Training script  
├── main.py           # Main entry for prediction & visualization  
├── preprocess.py     # Contour detection and filtering logic  
├── chopped.py        # Postprocessing (resizing, border fixing, etc.)  
├── dataset/          # Directory for /train and /test directories  
├── input/            # Directory for input images to be recognized  
├── output/           # Inference results with annotated predictions  
├── char_cnn.pth      # Trained model weights  

Usage

python main.py Ensure images of car are in ./input. Annotated results will be saved to ./output.

Acknowledgement

• Chars74K Dataset
• OpenCV & PyTorch