A Comparative Analysis of Various Deep Learning Models for Traffic Signs Recognition from the Perspective of Bangladesh

As they play a significant role in autonomous driving and traffic safety, traffic sign identification and recognition have recently emerged as one of the most significant fields in image processing and computer vision. Early studies in this field offered several deep learning-based methods for classifying distinct traffic signs using various standard datasets. However, not many researchers focused on creating a dataset of traffic signs in Bangladesh and applying deep learning techniques to recognize them. In this research, we compare and contrast several deep learning models for recognizing traffic signs from the perspective of Bangladesh. We construct a novel dataset with over 2000 images representing thirteen distinct kinds of typical traffic signs in Bangladesh. Using data augmentation, about 8386 images are generated from the original dataset. Subsequently, transfer learning and fine-tuning approaches are applied to nine different deep learning models using this dataset, and the outcomes are compared. Results indicate that ViT had the highest validation accuracy of 99.91% for fine-tuning, while DenseNet201 had the highest validation accuracy of 99.86% for transfer learning. Almost all models attained excellent training and validation accuracy levels, showing that they were able to successfully learn the dataset’s characteristics.