Bayesian Optimized Traffic Sign Recognition on Social Media Data Using Deep Learning

Abstract

One of the distinguishing features of traffic signs is the wide variation in their visual appearance in actual environments. Road signs are affected, for example, by variations in lighting, shifting weather patterns, and partial obstructions. Humans are very skilled at reading traffic signs, thus they are designed to be easy to read. However, identifying traffic signs still appears to be a difficult pattern recognition task for computer systems. This research concentrates on transfer learning using deep convolutional neural network (CNN) and its designs, including VGG16, ResNet101, and EfficientNetB2, to handle such challenges. We then adjusted the hyperparameters using Bayesian Optimization. The German Traffic Sign Recognition Benchmark Dataset (GTSRBD), the Chinese Traffic Sign Recognition Benchmark (CHTSRB), and the Indian Traffic Sign Dataset (ITSD) are the three datasets used in the study to train and test these pre-trained CNN classifiers without their TOP layers. According to experimental findings, the suggested approach performed well in terms of metrics for evaluating the identification of traffic signs. The ResNet101 model performs better than all other implemented models on all three datasets, it gives the accuracy of 97.6537% on GTSRBD, 96.7677% on CHTSD, and 90.9547% on ITSD.