Author(s)

Zhe Liu, Xiao Li and Jiaxian Yang

Corresponding Author

Zhe Liu

ABSTRACT

To improve the accuracy and real-time performance of traffic sign image recognition and reduce the computational cost, a traffic sign recognition algorithm based on an improved convolutional neural network was constructed. Firstly, the data set is pre-processed using image graying processing, limited contrast histogram equalization method and random erasing method to obtain high-quality data set. Then a traffic sign recognition model TSR_ConvNet is constructed. To reduce the occurrence of overfitting, the dropout layer is added and the Label-smoothing regularization method is used. Added the Batch Normalization layer to normalize input batch samples. And by analyzing the effect of the size of the convolution layer filter on the extracted feature map, the convolution kernel of the optimal size is selected. The German traffic sign data set (GTSRB) was used to perform traffic sign classification and recognition experiments. Experimental results show that a recognition accuracy rate of more than 98.74% and a recognition speed of 17 ms per image are obtained on the GTSRB benchmark data set. The traffic sign recognition model TSR_ConvNet constructed in this paper has the characteristics of short training time, good generalization ability, high recognition accuracy, and fast recognition speed.

KEYWORDS

Traffic sign recognition; convolutional neural network; deep learning; Image preprocessing; overfitting