An Enhanced Traffic Sign Recognition Approach Based on Federated Learning

Traffic sign recognition (TSR) plays a significant role in modern vehicle control systems and autonomous driving technologies, improving road safety and operational effectiveness. Numerous present approaches for TSR are either overly complex or lack regularization. This study presents a customized deep-learning method that combines image augmentation and K-fold cross-validation techniques to enhance generalization and overcome these limitations. Our straightforward federated learning global model consists of four convolutional layers and two dense layers, achieving an impressive average test accuracy of 98.73% and after k-fold cross-validation 99.09%. Moreover, bias-variance investigation validates the regularization of our model, ensuring consistent performance on the German Traffic Sign Recognition Benchmark (GTSRB) datasets. This promises gains in safety and operational reliability for intelligent transportation systems, indicating its efficiency and efficacy in recognizing traffic signs.