Traffic-Sign-Classification-using-Computer-Vision

📝 Problem Statement

The goal of this project is to classify German traffic signs into their correct categories using deep learning models.
Traffic sign recognition is a critical component of autonomous driving systems and advanced driver-assistance systems (ADAS), where real-time classification must be robust against lighting variations, rotations, occlusions, and class imbalance.

---

📊 Data Source

• Dataset: German Traffic Sign Recognition Benchmark (GTSRB)
• Access: Available directly via torchvision.datasets.GTSRB
• Content: 43 different traffic sign classes with thousands of labeled images
• Structure: Training, validation, and test splits with a stratified distribution

Example samples from the dataset:

---

🔄 Preprocessing

To prepare and balance the dataset for training:

• Image Processing:
  • Resize to 64×64 pixels
  • Convert to tensors (ToTensor)
  • Normalize pixel values
• Data Augmentation:
  • Random rotations (RandomRotation)
  • Affine transformations (RandomAffine)
  • Color jitter (ColorJitter)
• Class Imbalance Handling:
  • Stratified splits and balanced sampling ensured minority classes were not underrepresented.

---

🏗️ Models Used

Three architectures were implemented and compared:

1. Custom MLP Classifier

   • Fully connected layers (flattened input)
   • Dropout and BatchNorm1d for regularization
   • Serves as a baseline

2. Custom CNN Classifier

   • Multiple convolutional layers with ReLU, BatchNorm2d, and MaxPooling
   • Stronger regularization and capacity to capture spatial features
   • Designed specifically for 64×64 traffic sign images

3. ResNet-50 (Transfer Learning)

   • Pretrained on ImageNet via torchvision.models.resnet50
   • Final layer replaced for 43 GTSRB classes
   • Partial fine-tuning (only last block unfrozen)

---

✅ Results

Metric	CNN	MLP	ResNet-50
Accuracy	98.84	65.84	96.97
Precision (Macro)	98.32	64.90	95.14
Recall (Macro)	97.99	65.22	96.71
F1 (Macro)	97.98	63.50	95.79
Precision (Micro)	98.84	65.84	96.97
Recall (Micro)	98.84	65.84	96.97
F1 (Micro)	98.84	65.84	96.97

📌 Discussion

• Best Model: The custom CNN achieved the highest accuracy (98.84%), slightly outperforming ResNet-50 (96.97%).
• MLP Baseline: The MLP lagged far behind (65.84%) since it flattened images and ignored spatial patterns.
• Macro vs. Micro: Near-equal macro and micro scores for CNN/ResNet-50 indicate balanced performance across all classes, thanks to stratified splits, augmentation, and balanced sampling.
• CNN vs. ResNet-50: Despite transfer learning, the CNN slightly outperformed ResNet-50, likely due to:
  1. Partial fine-tuning of ResNet-50
  2. Domain gap between ImageNet and GTSRB
  3. CNN being compact and tailored to 64×64 inputs with stronger regularization

---

🚀 Conclusion

• Custom CNN proved to be the most effective, achieving nearly 99% accuracy, surpassing ResNet-50.
• Balanced sampling and targeted augmentation were key in mitigating class imbalance.
• The MLP baseline was useful for comparison but inadequate for fine-grained visual recognition tasks.
• Future work:
  • Fully fine-tune ResNet-50 with more training epochs
  • Explore more advanced architectures like EfficientNet or Vision Transformers