Research On Convolutional Neural Network For Compression Algorithm

In recent years,convolutional neural networks have been widely used in computer vision tasks,such as image recognition and object localization.Convolutional neural network compression technology can effectively reduce the amount of calculation and parameters of the neural network,so that the computing device can run the convolutional neural network model more smoothly,reduce device energy consumption,and improve the user experience.In this paper,the research of convolutional neural network compression algorithm is divided into the following two aspects.In the aspect of neural nodes pruning,combining the influence of convolutional layer and batch normalization layer on the output feature map of neural node,this paper propose a neural node importance degree evaluation algorithm to score the importance of neural nodes in the convolutional layer.Pruning a part of the neural nodes with the lowest importance scores to achieve the compression of the convolutional neural network;According to the distribution of the importance scores of the neural nodes in the convolutional layer,an algorithm for calculating the pruning rate of the neural nodes of each convolutional layer is proposed.If the importance scores of the nodes are similar,the pruning rate will be smaller.The experimental results on the VGG16 network trained by the ILSVRC2012 dataset are that 63.9% of the convolutional layer neural nodes are pruned,the number of floating-point calculations is reduced by 66.5%,and the Top-5 accuracy is reduced by 0.56%.In the aspect of storage space compression: A dictionary learning convolutional neural network storage space compression algorithm is proposed.The linear combination of a small number of entries in the dictionary approximates the weight parameters of each channel in the neural node,and the coefficients of the linear combination of entries Quantify.When storing convolutional neural network parameters,only the dictionary,the quantized coefficients,and the index of the entries in the dictionary corresponding to the coefficients need to be stored.The experimental results of the storage space compression of the VGG16 model are that the accuracy is reduced by 0.09%,when the storage space compression is 12.63% of the original model.