Traffic sign recognition algorithm based on depthwise separable convolutions

In order to solve the problem of low precision detection for small and medium-sized targets in the real-time traffic sign recognition, a lightweight traffic sign detection network based on YOLOv3-tiny depthwise separable convolution is proposed. In this paper, a deep separable convolution module is constructed by using deep separable convolution instead of the common convolution to construction feature extraction network, the feature information of small and medium-sized targets is better extracted under the premise of ensuring the calculation amount.At the same time, the multi-scale feature fusion network is improved to improve the detection accuracy of small and medium-sized traffic signs. The h-swish activation function is used to reduce the image features lost due to the increase of the number of network layers, and the detection of multiple types of traffic signs is realized.The experimental results show that the algorithm effectively improves the detection of small and medium-sized traffic signs. Warining, mandatory and prohibitory traffic signs are detected on the verification set. The detection accuracy (AP) is 98.57% , 96.03% and 98.04% respectively. The average detection accuracy (mAP) was 97.54% and the detection speed was 201.5 f/s. The average accuracy was 14.01% higher than that of YOLOv3-tiny. The algorithm effectively improves the detection accuracy under the premise of ensuring low calculation of light network and good timeliness of detection.