Application Of Convolution Neural Network In Traffic Image Recognition

The application of convolution neural network(CNN)in traffic image recognition is a hot research content of intelligent transportation system.The classification of traffic signs and vehicle types is the two core parts of traffic image recognition,which faces many difficulties in practical application.The method of manually designing feature extractor is slow and low accuracy,while convolution neural network is simpler and has higher classification accuracy.Therefore,this thesis studies the classification of traffic signs and vehicle images based on convolution neural network(1)This thesis summarizes the difficulties and main application scenarios of traffic sign and vehicle type recognition,and pays attention to the latest research progress at home and abroad;summarizes the development,theoretical basis,working principle,optimization algorithm and representative improved algorithm of convolutional neural network.(2)Traffic sign recognition.Traffic sign recognition will be disturbed by the complexity of natural scene,and the change of image size will also lead to the degradation of network performance.In order to improve the influence of image size on network performance,firstly,the histogram equalization method is used to enhance the features of traffic sign data,and then the enhanced image is normalized to three sizes:large,medium and small;secondly,the network structure and convolution kernel size are constructed to adapt to the image size for feature extraction,and the multi-scale features are fused in the full connection layer of the overall recognition network;Finally,traffic sign images are classified by extreme learning machine.(3)Vehicle type identification.Due to the uneven information distribution of the upper and lower parts of the car face,inspired by the attention mechanism,the car face is divided into two regions,and two different networks are designed for recognition.Firstly,the vehicle face image is divided into upper and lower parts by using the segmentation function on the vehicle recognition data set;secondly,two backbone networks are designed to extract vehicle features respectively.Because the lower part of the vehicle face is rich in information,the bilinear convolution neural network model is selected to extract the lower part of the vehicle face features,and the shallow and high-level features are fused to obtain the feature vector with more rich information;finally,the feature vector with more abundant information is obtained Then,before the full connection layer of the fusion network,two layers of 1 × 1 convolution kernel are used for feature superposition and dimension reduction to enhance the practicability of the network.Aiming at the problem of uneven distribution of traffic signs and the two parts of vehicle faces,the convolution neural network is improved.The improved methods are verified on GTSRB data sets and Compcars datasets respectively.All of them have achieved good classification accuracy,which proves that the improved algorithm is advanced.