Featured Projects

Traffic Sign Recognition

Summary

Designed and trained a CNN to classify traffic signs from raw images with reliable generalization on unseen data.

Problem

Build a system that classifies traffic signs from raw images despite real-world variability (lighting, angle, noise). The goal is reliable generalization on unseen examples, not just fitting training data.

Approach

Preprocessing (resize + normalize), CNN architecture (convolution + pooling + dense classifier), supervised training with validation monitoring and iterative tuning.

Highlights

• Preprocessing pipeline for consistent input shape and normalization.
• CNN with convolution, pooling, and dense classifier layers.
• Validation monitoring to guide iterative tuning and prevent overfitting.

Results

Test accuracy

(varies by run)

Model convergence observed across multiple runs.

Evaluation

Supervised learning with train/validation/test split

Baseline: Random classification baseline

Limitations

Sensitivity to extreme lighting conditions and rare classes.
Limited interpretability beyond activations (no explainability layer in v1).

In progress: adding benchmarks and visuals.

Trade-offs

• Model capacity vs generalization trade-off requires careful validation.
• Fixed test evaluation means the model must hold on unseen data.

Next improvements

• Data augmentation experiments to improve robustness.
• Regularization tuning (dropout and weight decay).
• Explainability layer (Grad-CAM or saliency maps).
• Class imbalance handling if needed.