A Comparative Analysis of CNN Models and Data Augmentation for Optimizing Traffic Sign Recognition

Abstract

Correctly and efficiently recognizing traffic signs plays a crucial role in autonomous driving. his capability ensures that self-driving cars can obey traffic rules, respond to signals in real-time, and make appropriate decisions to avoid accidents and ensure the safety of all road users. However, the highly complex real-world settings still make this task challenging. This research aims to explore and evaluate multiple models and data augmentation strategies to find the optimal balance between performance and computational cost. Specifically, this research utilized the Traffic Sign Recognition Benchmark dataset and compared ResNet50, MobileNetV2, and a self-defined convolutional neural network model to evaluate their test accuracy and the total number of parameters used. By analyzing the results, the ResNet50 model achieved the highest accuracy. Meanwhile, MobileNetV2 demonstrated greater robustness, generalizability, and lower computational cost, while still maintaining a reliably high accuracy. Among all the data augmentation methods tested, adjusting brightness and contrast outperformed the image data generator and color jitter in almost all models and input dimensions.