Research On Image Classification Based On Lightweight Convolutional Neural Network

With the continuous development of computer vision,convolutional neural network is making progress in the field of image classification.Using convolutional neural network to classify images is more efficient and convenient than traditional image processing methods.With the continuous progress of convolutional neural network,researchers put forward a lot of convolutional neural network models with more complex structure and deeper layers,but the following It is efficiency problem.Although convolutional neural network with complex structure will have better accuracy and other advantages,it is a short board that can not be ignored because of its low efficiency,large amount of parameters and high amount of calculation.Therefore,lightweight convolutional neural network is proposed to improve the efficiency of image data classification while ensuring the accuracy and reduce the amount of parameters,which can be used in small terminal equipment The research of lightweight convolutional neural network has far-reaching significance and influence.The main research results are as follows:(1)The algorithm and principle of image classification based on deep learning are fully studied,including the algorithm principle,algorithm process and activation function of neural network in the process of image classification.Then,the traditional three classic convolutional neural networks,Alex Net,Vgg Net and Google Net,are studied.According to the analysis of the classification accuracy of the three networks on Image Net2012,the parameter amount and calculation amount of each network,the analysis results show that although the classification accuracy of the network is high,the parameter amount and calculation amount of the three deep networks are high,that is,the network efficiency The rate is low,so the next step is to study the lightweight network based on image classification.(2)Based on the research of Squeeze Net,this paper proposes a Vans Net network based on Squeeze Net,which is a lightweight convolutional neural network.Aiming at the problem that the accuracy of Squeeze Net is not ideal due to the compressed parameters,this network proposes a scheme to introduce residual network to increase the hop layer structure and broaden the network width.Then,Vansnet,Alex Net,Res Net and the original network are combined The experimental results show that the image classification accuracy of the improved Vans Net is higher than that of the Squeeze Net in the two data sets,and better than that of the Alex Net and Res Net in lightweight.(3)This paper puts forward the application of group information fusion strategy,and studies the group convolution and feature fusion fully.The idea of group convolution is intended to reduce the amount of calculation and parameters to improve the efficiency,followed by the problem of information non circulation between groups after grouping.Therefore,the idea of feature fusion is introduced to solve this problem,and the idea of batch normalization is introduced to improve the network efficiency to design Xans Net In lightweight network,firstly setdifferent packets to determine the number of packets of Xans Net,then put forward the design scheme of Xans Net and complete the design,then compare and analyze the image classification experiment of Xans Net and other four network structures in two different data sets,finally according to the analysis of the experimental results,the experimental results show that the accuracy of Xans Net classification has a good performance.