Traffic Sign Recognition

A deep learning model for recognizing and classifying traffic signs using PyTorch. This project demonstrates training a ResNet18 model on the German Traffic Sign Recognition Benchmark (GTSRB) dataset.

📊 Model Performance

• Test Accuracy: 93.35%
• Model Architecture: ResNet18 with custom classification head
• Input Size: 224x224 pixels
• Number of Classes: 43

Detailed Metrics

Metric	Score
Precision (macro avg)	0.88
Recall (macro avg)	0.87
F1-score (macro avg)	0.87
Precision (weighted avg)	0.94
Recall (weighted avg)	0.93
F1-score (weighted avg)	0.93

🚀 Features

• Data preprocessing and augmentation
• Model training with early stopping
• Learning rate scheduling
• Model evaluation with detailed metrics
• Confusion matrix visualization

🛠️ Project Structure

.
├── data/
│   └── processed/          # Processed dataset
│       ├── images/         # Extracted images
│       │   ├── train/      # Training images
│       │   ├── valid/      # Validation images
│       │   └── test/       # Test images
│       ├── labels.csv      # Image paths and labels
│       └── label_names.csv # Class names
├── src/
│   ├── train_fixed.py      # Training script
│   ├── evaluate.py         # Model evaluation
│   └── process_pickle_data.py # Data preprocessing
├── best_model.pth         # Best model weights
├── final_model.pth        # Final trained model
└── requirements.txt       # Python dependencies

🚀 Quick Start

1. Clone the repository

   git clone <repository-url>
   cd traffic-sign-recognition

2. Set up a virtual environment

   python -m venv venv
   source venv/bin/activate  # On Windows: venv\Scripts\activate

3. Install dependencies

   pip install -r requirements.txt

4. Train the model

   python src/train_fixed.py

5. Evaluate the model

   python src/evaluate.py

📊 Results

After training, you'll find:

• Training/validation loss and accuracy plots
• Confusion matrix saved as confusion_matrix.png
• Detailed classification report in the console

📝 Notes

• The model was trained for 5 epochs with early stopping
• Data augmentation includes random rotations, flips, and color jitter
• The best model is saved as best_model.pth based on validation accuracy

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.