Remote Sensing Image Scene Classification Based On Attention Residual Network And Soft Thresholding Learning

With the development of remote sensing technology,its application has been increasingly widespread.Remote sensing image scene classification,as a significant application of remote sensing technique,has always become a research highlight of experts and scholars.At present,with the dramatic increase in the number of remote sensing images,more and more high-resolution remote sensing image data collections appear.The traditional scene classification methods based on feature and semantics can no longer meet the application requirements.In recent years,data-driven in-depth learning has developed rapidly.With the convolution neural network,deep learning-based classification methods can easily extract the complex and rich high-level features in images,and have become the main method of remote sensing image scene classification.However,most of them have significant effects only on the task of remote sensing image scene classification under specific small-scale datasets.For the classification difficulties of remote sensing image scene classification under a certain scale data set,such as complex background,large difference of target size,high similarity of different scenes and large difference of similar scenes,most existing classification methods cannot effectively extract image features,the accuracy of classification needs to be improved,and the generalization ability is poor.Therefore,better methods are needed to further improve the accuracy.In order to solve the problem that most of the existing classification methods do not perform well in a certain scale of remote sensing image data collection,this thesis studies the principle of the main convolution neural network and attention mechanism,puts forward a classification method based on attention residual network and soft threshold learning,improves the Bottleneck module structure on the basis of residual network Res Net50,adds channel attention mechanism and soft threshold operation.Attention Residual Network ARN and Attention Residual Filtering Network ARFN are designed and applied to scene classification of remote sensing images.Finally,a large number of experiments on three datasets show that the proposed remote sensing image scene classification method can effectively improve the classification accuracy under a certain scale of datasets.To be exact,the main research contents and achievements this thesis includes the following:（1）Remote sensing image scene classification based on attention residual network.Considering the problem of large intra-class differences and high similarity between classes in scene classification of remote sensing images,the theoretical principles and classical algorithms of attention mechanism are studied,and a classification method based on attention residual network is proposed.Based on this method,an attention residual network is designed.Channel attention mechanism is introduced into the residual network,and the potential correlation information between objects in the scene is used to optimize the extraction process of remote sensing image feature information,so that the attention residual network can focus on the local area with the richest feature information in the classification process of remote sensing image scene,and filter out the key feature information needed by the task,so as to improve the classification accuracy.（2）Remote sensing image scene classification based on attention residual network and soft threshold learning.Considering the problem of complex background and large difference of target size in remote sensing image scene classification,the soft threshold function in in the field of signal noise reduction is studied,and a classification method based on attention residual network and soft threshold learning is proposed.This method improves ARN by combining the soft threshold function in the signal noise reduction algorithm,introduces the soft threshold learning module,and designs the attention residual filter network.Attention residuals filter network can use channel attention mechanism to update learning threshold,and then use soft threshold function to filter out unwanted noise redundant information in classification tasks,so as to improve classification accuracy.