Traffic Sign Recognition Based on Multi-branch Residual Mixture Model

In order to improve the accuracy of traffic sign recognition in real environment, a multi-branch residual mixture model based on LeNet-5 is proposed. Firstly, erosion and dilation in morphology are used to preprocess images. Secondly, two relatively independent convolutional branches are constructed to extract the multi-scale feature information of the images. Based on the idea of residual connection, bottleneck blocks are constructed to increase the depth of the network to improve the fitting ability of the model. In addition, Squeeze-and-Excitation block is added to the improved algorithm to reinforce the useful features, and Dropout is added to alleviate the overfitting. Finally, the hyperparameters are selected by using the five-fold cross-validation, and the exponential decay learning rate is used instead of the original fixed learning rate, so that the model can converge better. The accuracy of the improved algorithm on GTSRB and BelgiumTSC is better than that of the classical LeNet-5 model.