Empirical Analysis of Traffic Sign Recognition using ResNet Architectures

Traffic sign recognition is an important task in the field of computer vision and is often used in various applications such as autonomous vehicles, intelligent transportation systems, and road safety systems. Identifying and categorizing the many traffic signs that are present in the environment, such as speed limits, stop signs, and pedestrian crossings, among others, is the goal of traffic sign identification. In this study, we use all iterations of the well-known deep neural network architecture such as Residual Neural Network (ResNet) to conduct an empirical analysis of the recognition of traffic signs. Using the recognition benchmark dataset of German Traffic Signs (GTSRB), which contains 43 different classes of traffic signs, we assess the performance of different ResNet architectures such as ResNet50, ResNet50V2, ResNet101, ResNet101V2, ResNet152 and ResNet152V2. We evaluate all these ResNet architecture performances based on their accuracy, precision, recall, and F1 score. According to the outcomes of our experiments, ResNet50V2 achieves the maximum accuracy of 99.78%, while ResNet152 achieves the lowest accuracy of 99.49%. We also noted that the accuracy of the ResNet50V2 model grows along with the architecture’s depth. To enhance the performance of these DL architectures, image preprocessing techniques such as resizing, zero-centring and normalization is applied to the traffic sign images. The reshaped matrix of the pre-processed images is then fed into the ResNet models according to their architectures. The results of this study demonstrate that the power of different ResNet architectures with image pre-processing, reshaping the matrix, and training time can significantly improve the performance of traffic sign recognition.