Research On The Models For Remote Sensing Image Scene Recognition Based Convolutional Neural Networks

Remote sensing image scene recognition can explain the semantic content of remote sensing images and contribute to complete remote sensing task like image classification,target detection.In recent years,deep convolutional neural networks have been widely used in image classification,image target recognition and other fields and have achieved excellent consequent.The convolutional neural network can extract high-order,abstract features of images,which is needed in remote sensing image scene recognition tasks.Therefore,convolutional neural networks have been used in remote sensing image scene recognition tasks.In view of the situation that the deep neural network model consumes a large amount of memory and computing resources,in this thesis,deep compression has been used for the scene recognition model of remote sensing images,after the network pruning,weight sharing and quantization,the model size has been compressed without affecting the performance of the model on a large scale.Firstly,the basic structure and training process of the convolutional neural network have been introduced,the network structure and important parameters used in the experiment have been selected according to the actual situation.By studying different number of convolution layers,different number of convolution kernels,different initial learning rates and different learning rate update frequencies for the impact of the final recognition accuracy,the structure of the network and the value of important parameters have been determined,for AID datasets,the initial accuracy of the network used in this thesis is 96.2%,and then the activation function ReLu of the convolutional neural network is replaced by PReLu,and the recognition accuracy of the network is further improved to 97.1%.After the training is completed,the remote sensing scene images recognition model based the convolutional neural network has been obtained.Finally,the layers of the model have visually analyzed.The powerful performance of the convolutional neural networks relies on millions of parameters and complex network structures,which make the convolutional neural networks consume much of memory and computing resources during the training process,in view of the situation,in this thesis the method of deep compression has been used for the model: including network pruning,weight sharing and quantization.It’s necessary to analyze the weight value of the deep neural network model of remote sensing image scene recognition,then pruning the connections by the proposed threshold selection method,and the deredundancy of the network is realized by removing the unimportant connections.The improved K-Means algorithm has been used to achieve weights sharing and quantization by clustering the remaining connections.After the above two steps,the recognition accuracy of the model decreased from 97.1% to 95%,and the size of the model was compressed from 240 M to 20 M.