Efficient convNets for fast traffic sign recognition

Abstract

While deep convolutional networks gain overwhelming accuracy for computer vision, they are also well-known for their high computation costs and memory demands. Given limited resources, they are difficult to apply. As a consequence, it is beneficial to investigate small, lightweight, accurate deep convolutional neural networks (ConvNets) that are better suited for resource-limited electronic devices. This study presents qNet and sqNet, two small and efficient ConvNets for fast traffic sign recognition using uniform macro-architecture and depth-wise separable convolution. The qNet is designed with fewer parameters for even better accuracy. It possesses only 0.29M parameters (0.6 of one of the smallest models), while achieving a better accuracy of 99.4% on the German Traffic Sign Recognition Benchmark (GTSRB). The resulting sqNet possesses only 0.045M parameters (almost 0.1 of one of the smallest models) and 7.01M multiply-add computations (reducing computations to 30% of one of the smallest models), while keeping an accuracy of 99% on the benchmark. The experimental results on the GTSRB demonstrate that authors’ networks are more efficient in using parameters and computations.