Research On The Structure Design Method Of Convolutional Neural Network In Image Classification Task

In recent years,image classification and recognition technologies are increasingly used in medical imaging diagnosis,public security and driverless fields,and in some aspects the accuracy of classification has exceeded that of human experts.Compared with the traditional image classification methods,convolutional neural networks have shown excellent performance in image classification tasks.In order to pursue a higher classification accuracy,researchers have tried to deepen,broaden and optimize the architecture of neural networks,and propose new training methods.In the real world,some practical applications are equipped with platforms that have limited storage space and computing power,such as mobile devices.Thus,the neural network with a small amount of parameters and computations is more feasible in engineering.This paper studies the structure design of convolutional neural networks for the task of image classification.The main contributions of this paper are as follows.(1)To address the problem of losing "useful" information during max-pooling,the fully sampling method and the downsampling method based on L1 norm are proposed.The fully sampling method uses all inputs.The downsampling method based on L1 norm divides the input into 2?2?d(height ?width?depth)non-overlapping tensors(multidimensional arrays),and selects tensors with the same position to form four different new tensors.Then selects k(k = 1,2,3)tensors with the largest L1 norm from the four new generated tensors,and concatenate them in the direction of the channels to achieve downsampling in the dimension of height and width.In order to verify the validity of the proposed method,the experiments on CIFAR-10 dataset and MNIST dataset show that compared with max-pooling,the proposed method can improve the classification accuracy without increasing the parameter quantity.(2)In order to reduce the parameter quantity of convolutional neural networks,we study the model compression method based on structure design,and analyze the model compression efficiency of depthwise seperable convolution and group convolution.We also introduce the group shuffle method which is used to enhance the approximating ability of neural networks that use group convolution.The convolutional neural network constructed by the full sampling method and the down sampling method based on the L1 norm is compressed.The parameter amount and the classification accuracy can be weighed according to the actual situation.