Performance Analysis of CNNS for Traffic Sign Recognition Benchmark

Abstract

In this paper, a comparative study of convolutional neural network (CNN) architectural choices for traffic sign recognition is presented. The effects of different network depths, convolutional kernel sizes, pooling strategies, and activation functions on traffic sign classification performance are examined. In practical scenes, traffic sign images are often affected by illumination changes, scale variation, viewpoint variation, partial occlusion, motion blur, and complex backgrounds. These conditions make accurate traffic sign classification difficult. Therefore, CNNs need to extract discriminative visual features. Experiments are conducted on the German Traffic Sign Recognition Benchmark (GTSRB). The number of learnable parameters is used to analyse model complexity, and a row-normalised confusion matrix is used to examine class-level classification behaviour. Results show that the CNN with three convolutional layers, 4 × 4 convolutional kernels, max pooling, and ReLU activation achieves the best classification performance among the tested configurations. Among the examined architectural choices, network depth shows the largest difference in classification performance, followed by pooling strategy, convolutional kernel size, and activation function.