Research And Implementation Of Optimized CNN Network Model In Scene Classification Algorithm For High-resolution Remote Sensing Images

With the development of remote sensing satellite technology,remote sensing images have entered the era of high spatial resolution.High-resolution remote sensing images can provide related spatial orientation information,general features,and overall texture.It has a wide range of applications in image target recognition,monitoring and tracking.For a large number of high-resolution remote sensing images,it becomes particularly important to how to classify and archive useful information from them.The technology involved is scene classification.Based on the advantages of convolutional neural networks in the field of image classification,this paper proposes three improved models CNN-1,CNN-2,and CNN-3 to conduct classification research on the high-resolution remote sensing scene image data.And the compression of the proposed network model is accelerated.The network model proposed in this paper is composed of a spatial transformation network and a bilinear network.The spatial transformation network introduces an attention mechanism,which can obtain the area of interest of the image to capture the significant information of the main object in the image data.The linear network is a parallel network composed of two residual networks with improved residual units,One of the networks is used to extract the object features of the image,and the other is used to extract the position features of the image.The fusion of the features extracted by the two networks can improve the classification accuracy of the network model.The spatial transformation network and the bilinear network cooperate with each other to complete the task of scene classification.The aim is to effectively improve the classification performance of network models for high-resolution remote sensing scene images.At the same time,in order to solve the problem of"small differences between classes and large differences within classes"in high-resolution remote sensing images,this paper proposes a new loss function（combination of Softmax-Loss and Center-Loss）.Based on the further analysis,it is found that the network model proposed in this paper has a larger volume,a longer training time,and more redundant parameters,which affects the transplantation of the model.Therefore,the weighted pruning algorithm and dynamic weighted pruning algorithm are used to optimize the network model.In this paper,the AID dataset and the NWPU-RESISC45 dataset are selected as the experimental data in this paper.The experimental results show that:on the AID dataset,the classification accuracy of the CNN-1 model,CNN-2 model,and CNN-3 model are 97.68%,98.56%,98.82%,which are 2.94%,3.82%,4.08%higher than the VGG-VD16^[47]model respectively.On the NWPU-RESISC45 dataset,the accuracy rates of the three models are94.81%,95.12%,95.65%,which are 3.51%,3.82%,4.35%higher than the IORN4-VGG16^[48]model respectively.The comparative experimental results of the weighted pruning algorithm and the dynamic weighted pruning algorithm show that the dynamic weighted pruning algorithm has a higher speedup ratio and a lower accuracy loss value when the compression effect is similar.Finally,gaussian white noise with different intensities is added to the test sample set for experiment.The result proves that the network model after compression and acceleration can obtain better robustness.