As part of an academic deep learning project, I designed and evaluated multiple convolutional neural network (CNN) architectures to classify traffic signs using the German Traffic Sign Recognition Benchmark (GTSRB) — a real-world dataset containing over 50,000 labeled images across 43 traffic sign classes.

The project involved implementing and comparing three architectures:

1. Custom CNN:
A compact convolutional neural network designed from scratch using PyTorch. It included two convolutional layers, two fully connected layers (512 → 43 outputs), and a dropout layer for regularization. Despite its simplicity, the model achieved 98%+ accuracy in just 10 epochs.

2. ResNet50 (Transfer Learning):
I fine-tuned the final block of a pre-trained ResNet50 model and replaced its fully connected layer to adapt it to the GTSRB classification task. After 20 epochs, it achieved 99.76% accuracy with minimal loss (0.009), showing strong generalization and optimization.

3. VGG16 (Transfer Learning):
Fine-tuned all layers of the VGG16 architecture, again replacing the final layer for 43-class output. This model yielded the highest performance with 99.86% accuracy and a validation loss of just 0.0043 after 20 epochs.

All models were trained using:
• PyTorch
• Batch size = 32
• Learning rate = 0.001
• Weight decay = 1e-5

Data preprocessing included image normalization, resizing, rotation, and flipping (data augmentation), ensuring robust generalization across varied visual conditions.

Throughout the project, I conducted hyperparameter tuning, performance visualization across epochs (loss/accuracy curves), and qualitative evaluation through sample predictions. All models correctly classified the same set of test images, demonstrating their reliability.

This project deepened my expertise in CNN architecture design, transfer learning, model evaluation, and hands-on PyTorch implementation, while reinforcing best practices in reproducibility and training efficiency.