An Image Classification Method Based On An Improved GoogLeNet Network Model

With the rapid development of neural networks and image processing technology,the methods of applying neural networks to image processing have become increasingly mature,and convolutional neural networks have more and more applications in daily life.This article focuses on the problem of low classification accuracy of convolutional neural network GoogLeNet in remote sensing images.Onthe basis of GoogLeNet,deep separable convolution and dense connections in De-nSeNet are introduced.The main research content of this article is:1.Introduce the basic structure of classical convolutional neural networks.From the perspective of mathematical formula,explain why activation function should be used and what impact the use of different activation function will have on network training.2.Research and improve on the problem of convolutional neural network GoogLeNet not being able to fully extract multiscale features from remote sensing images.7X7 convolution channel and hole convolution are introduced on the basisof the original Ince-ption module.Then,the residual module in ResNet network is introduced.In terms of the activation function,the H-Swish activation function,which is faster in calculation,is used.The DRGoogLeNet network and the originalGoogLeNet network were tested on the UCM dataset,and the accuracy of DRGoogLeNet was improved by 1.11%compared to the original GoogLeNet network.3.In order to address some potential drawbacks of the above improvements,further improvements are been proposed to the GoogLeNet network model.Firstly,hollow convolution is replaced by deep separable convolution with ensures so that further reducing the number of network parameters while all features can be fully utilized;then replace the residual module based on the Inceptionmodule with a dense connection module to ensure the full utilization of information features in network transmission.Finally,the DMGoogLeNet network model,DRGoogLeNet network model,and the original GoogLeNet network model were trained and testedon the UCM dataset and RSSCN7 dataset,respecti-vely.On the UCM dataset,the DMGoogLeNet network improved its accuracy by 1.83%compared to the original network,and on the RSSCN77 dataset,the DMGoogLeNet net-work improved its accuracy by 5.76%compared to the original network.