This project trains a convolutional neural network to classify real-world traffic signs under noisy conditions, demonstrating applied computer vision fundamentals and disciplined model evaluation.
Project Links
As a capstone to Harvard’s CS50AI, this project revisited a classic computer vision task using a more challenging dataset than MNIST's handwritten numbers. It marks a shift from tinkering to building models thoughtfully, learning from mistakes, and understanding how they handle real-world challenges. And it was a fitting note on which to bid a fond farewell to a series of pedagogically perfect courses that helped build my computer science foundation.
Project Overview
This project explores convolutional neural networks (CNNs) and computer vision using the German Traffic Sign Recognition Benchmark (GTSRB). Building on Harvard's CS50AI coursework, I developed and compared several model architectures, examining how design choices affect classification performance.
A compact CNN with two convolutional layers and batch normalization performed best, achieving 99.4% validation accuracy and approximately 99% on the held-out test set. More complex models performed worse, while class weighting improved recognition of less common signs at the expense of overall accuracy.
Error analysis showed that most remaining misclassifications involved visually similar speed-limit signs. The project combines systematic model experimentation, evaluation, and visualization to explore the trade-offs between model complexity and performance.
Gallery
CNN architecture: mimics the human eye by combining simple
features into complex object recognition.1
Class imbalance: Label distribution is highly skewed across
classes.
Dataset diversity: representative image for each label.
Validation and training curves: best model vs. baseline.
Confusion matrices: best model noticeably reduces off-diagonal
errors.
Misclassification confidence histogram: overkill model’s errors
are less confident and more frequent.
References
Dataset provided by the German Traffic Sign Recognition Benchmark (GTSRB).
1 Inspired by Deep learning for wireless capsule endoscopy: a systematic review and meta-analysis, Soffer, Shelly, et al, Gastrointestinal Endoscopy, © 2020 American Society for Gastrointestinal Endoscopy. Image generated with AI and edited by Bryan Johns. Used for educational purposes only. If this use does not qualify as fair use, I am happy to remove the image upon request.