For the traditional lenet-5 convolution neural network used in traffic signs and other kinds of recognition tasks, the problems such as low recognition accuracy, easy network over-fitting and gradient disappearance are improved. Inception convolution module group was cited to extract rich features of the target while increasing the depth of the network,Introduce the BN layer to normalize the input batch samples to improve the input of the neural network,At the same time, the better Relu activation function was used, and the global pooling layer was used instead of the full connection layer, and the size and number of convolution kernels were reasonably changed.The research results show that the improved LeNet-5 network can effectively solve the problems of over-fitting and gradient disappearance, and has better robustness.At the same time, compared with CNN+SVM and traditional lenet-5 network, the accuracy of the improved network classification can be up to 98.5%, which is 5% higher than CNN+SVM and 3% higher than traditional lenet-5 network, The accuracy of image recognition is improved significantly.