Siamese Attention Fusion Network Based Change Detection Of Remote Sensing Image

With the continuous development of deep learning theory,the application researches of deep learning in change detection of remote sensing image are also deeper and deeper.The change detection methods based on deep learning is driven by data to realize the characterization learning of different time series images,highlight the changes of interest,and reduce the effect of those irrelevant changes caused by external interference factors.Starting from the deep attention mechanism and Siamese network architecture,this thesis proposes three different deep fusion network based change detection algorithms.The main works of this thesis are as follows:(1)A Siamese fusion network model based on SE attention mechanism(SE fusion module)and a remote sensing image change detection algorithm based on this model(SE-SiamResnet)are proposed.The SE fusion module can establish the connection between the two branches of the Siamese network,so that it can adaptively assign weights to these channels of each feature image during the training process to enhanc the model’s learning ability of the changed and unchanged features between the dual-phase images.SE-Siam-Resnet embeds multiple SE fusion modules into the Siamese change detection framework and trains the network with certain label data.The difference map is generated automatically using this network and the final change detection reslult is obtained by using a simple threshold segmentation method.(2)A remote sensing image change detection algorithm based on the global second-order pooling model(Global Second-order Pooling,GSo P)and twin fusion network is proposed.In this method,GSo P based on the second-order statistics is used to calculate the attention weight of each channel.The fusion and intercommunication of the information of the two network branches is realized,and three fusion models with different structures are designed.Model 1 stitches the feature images of the twin network together along the channel dimension,and uses its covariance matrix to express the correlation between the channels.Model 2 merges the two feature images of the twin network,the feature image after fusion has the same number of channels as before the fusion,and its covariance matrix contains richer correlation information.Model 3 separately calculates the covariance matrix of the two feature images,uses grouped convolution to model the covariance matrix that is stitched together,and adaptively assigns weights to each channel.It can be seen from the experiments in this thesis that these three twin network fusion models have effectively improved the performance of the network.(3)A remote sensing image change detection algorithm based on twin UNet fusion(SE＿UNet)is proposed.SE＿UNet uses the UNet structure to provide richer neighborhood information for each pixel,and at the same time uses the SE attention mechanism to fuse the two feature images generated by the twin network,and adaptively assign weights to each channel of the feature image,which represents the changed feature The feature that is assigned a larger weight means that the feature that is not of interest is assigned a smaller weight.Experiments show that this method can improve the detection performance to a certain extent.