Research For Convolutional Neural Network Model Compression Algorithm Based On Object Detection

In recent years,the convolutional neural network(CNN)has become the main method in the field of computer vision,which has made a breakthrough compared with the traditional algorithm.However,convolutional neural network often has a huge amount of parameters and computation,which limits its application in embedded and other Embedded system.In this paper,aiming at the problem of large amount of computation and parameter redundancy of convolutional neural network,the faster RCNN(faster RCNN)algorithm is taken as the research object to compress and accelerate,and a pruning method based on statistical clustering is proposed.At the same time,the model is compressed with tensor decomposition method,and the influence of the selection of rank value on the hybrid compression effect is analyzed(1)Aiming at the improvement of detection accuracy of faster RCNN algorithm in Pascal VOC data set,in order to improve the recall rate of category objects with large difference in target size,this paper proposes a multi-scale region generation network,which enhances the foreground feature information in each anchor frame,improves the quality of candidate frame generation and detection accuracy of the algorithm.(2)In order to solve the problem of large computation and redundant parameters of convolutional neural network,this paper studies the structural pruning method based on statistics clustering.For pruning,a key but unsolved problem is the "relative importance" evaluation of parameters,which has a decisive impact on the performance of the model after pruning.Firstly,the distribution of convolution kernels in each layer of faster RCNN model on mean and standard deviation is analyzed.The "contribution degree" of each convolution kernel is divided into two categories by clustering.Secondly,the statistics and size of each convolution kernel are compared,and the convolution kernels with smaller contribution degree are cut during pruning.(3)In the network pruning method,some important convolution kernels may be pruned.A hybrid compression method of pruning method combined with tensordecomposition is proposed.The pruning method is adopted in the low-dimensional convolution layer,and the high-dimensional convolution layer is decomposed into three cascaded convolution layers with low rank approximation.The experimental results show that the pruning method combined with tensor decomposition can effectively reduce the parameters and calculation amount of faster RCNN convolution layer,improve the accuracy of the model after pruning alone.The compression rate of convolution layer parameters is 54% and the accuracy of Pascal VOC test set remains unchanged.At the same time,the forward computing speed of raspberry pi 4B system is 1.4 times.