Publication

Abstract : This study addresses the optimization of deep learning and transfer learning models for Traffic Sign Recognition (TSR) under diverse environmental conditions and class imbalances. Traffic Sign Recognition Database (TSRD) is used in this research study. Despite the widespread application of transfer learning in various domains, its implementation in TSR has been relatively unexplored. This research evaluates the efficacy of transfer learning in TSR using three neural network architectures: ResNet-18, ResNet34, and VGG-19. The study’s findings reveal that ResNet-18 outperforms others, achieving an accuracy of 88.47%, precision of 85.20%, and recall of 80.55%, while requiring only 254.136 seconds for training on the A100 GPUs. ResNet-34 also demonstrates considerable efficiency with an accuracy of 87.47%, but at a longer training duration of 372.95286 seconds. In contrast, VGG-19 shows limited suitability for TSR tasks, with a much lower accuracy of 5.32% and the longest training time of 1256.9595 seconds. The results underscore the potential of ResNet-18 and ResNet-34 in TSR applications, highlighting the need for a balanced approach towards accuracy, computational efficiency, and robustness in varying conditions. These insights are significant for advancements in autonomous navigation systems and enhancingroad safety.