Comparative Analysis of Deep Learning Models on GTSRB Dataset for Classifying Under-Represented Data

It is extremely difficult to address the intricacies of traffic sign recognition in driver assistance systems and self-driving cars. The creation of trustworthy and extremely accurate algorithms are essential for the broad integration of the recognition of traffic signs in various real-world contexts. However, poor traffic image quality makes this task difficult.impacted by elements like opacity, camera movement, bad weather coupled with inadequate lighting. This research aims to improve techniques for recognizing traffic signs by comparing four deep learning models' performance—Convolutional Neural Network (CNN), Vision Transformer (ViT), Swin Transformer, and Data-Efficient Image Transformer (DeiT)—on the German Traffic Sign Recognition Benchmark (GTSRB) dataset. An epoch count of minimum 10 was uniformly applied during the training phase for all models. The experimental results demonstrated that only the CNN and DeiT models were able to classify traffic signs correctly, while the ViT and Swin models failed to achieve accurate classification. This comparative analysis provides valuable insights for future research and development in the field of traffic sign recognition, emphasizing the significance of model selection in achieving optimal performance.