Type Identification Of Railway Freight Cars Based On Deep Learning

Traditional identification techniques of railway truck types exist following disadvantages:poor adaptability to the environment,sensitivity to light illumination,etc.The identification results are often decreased and the reliability is low.Applying the deep learning techniques to railway vehicle identification can reduce the image quality requirements of railway truck monitoring,and can improve the accuracy of automatic identification.The deep learning can further reduce the false detection rate and miss detection rate of railway vehicle identification,and can improve the robustness and reliability of railway vehicle identification.Baed on the theory and techiniques of deep learning and combined with three convolutional neural networks for object identification,this thesis constructs the identification models for railway vehicle,improve the structure of the networks,feature extraction and identification.The main contributions of this thesis are as follows:(1)To solve the problem of high complexity of the network,a method of network simplification is given.Baed on VGGNet-16 network,the number of fully connected layers is reduced,the nature and size of the pooled layer are changed to enhance the classification effect of the railway vehicle model.Experimental results show that the proposed method can preserve the feature extraction ability of networks,and reduce the network training time by one-third compared with the original network.(2)A classification identification method based on GoogLeNet and DS probability optimization is given.Firstly,the probability vector of the three classifiers outputted by GoogLeNet is judged by DS-based probability optimization.Secondly,the probability vector after DS probability optimization is used to characterize.Convergence improves the recognition accuracy of models based on the Goog LeNet network.Experimental results show tliat the proposed method can further improve the accuracy of the network by 5%of Top-1.(3)An improved residual network optimization method is given to reduce the computational complexity of the network and improve the network feature extraction ability and recognition accuracy.The method decomposes the standard volume integration in the network into deep convolution and point convolution,wherein the depth convolution is filtered by using different convolution kernels for each input charnel of the feature map,and the point convolution is used for the feature map.The feature information between different channels is fused,and the training model is compressed by decomposing convolution to reduce the calculation parameters.The experimental data proves that the improved method increases the recognition accuracy of Top-1 by 10%and the calculation by a quarter.