Research On Vehicle Classification Algorithm Based On Improved Convolutional Neural Network

With the continuous development of urban road traffic,people have an urgent need for intelligent transportation.As an important branch of intelligent transportation system,vehicle classification has attracted great attention.Traditional methods of vehicle classification have been difficult to meet the needs.This thesis focuses on the research of the distribution of training data,network structure and model compression around the vehicle classification problem of convolutional neural network in road traffic scene.This thesis mainly studies the following aspects:(1)The enhanced BIT-Vehicle dataset was used to respectively test AlexNet,VGGNet,GoogLeNet,ResNet and Densenet-BC networks of three different depths.After comparison,the performance of Densenet-BC was better than other models,and the classification accuracy of DenseNet-169 reached 94.83%.(2)To solve the problem that the DenseNet-169 network has a larger difference in the classification accuracy of different vehicle type on the dataset BIT-Vehicle,a model training method based on the adaptive reconstruction of data distribution is proposed.While reducing the difference in the classification accuracy of various vehicle type,This method increases the accuracy to 95.07%,and alleviate the problem of network convergence being too dispersed caused by data imbalance.(3)To solve the problem that the convergence speed of DenseNet-169 network is slow and the similarity of some vehicle type which makes it difficult for model to distinguish is high,a method of vehicle classification based on the two-channel DenseNet-BC network is proposed,which greatly improves the convergence speed of the network and improves the classification accuracy to 95.27%.(4)To solve the problem of containing excessive redundancy and useless feature information in the vehicle feature extracted by DenseNet-169 network,the redundant convolution kernel and useless convolution kernel in the model were removed with the method of model pruning.This method reduced the number of network parameters by 4.3M,which accounted for 30.39% of the number of parameters in the unpruned model.Meanwhile,the model test time was shortened by 98 ms on average,which accounted for 46.01% of the test time of unpruned model.In addition,the classification accuracy of the model after pruning was 94.93%The above work has been experimentally verified by the Tensorflow deep learning framework.Improved methods improves the performance of vehicle classification of DenseNet-BC network and can be used to solve the vehicle classification problem based on convolutional neural network.