A SAR Image Change Detection Method Based On Semantic Convolutional Neural Network And Generative Adversarial Learning

Synthetic Aperture Radar(SAR)is widely used in the field of remote sensing technology because of its all-time,all-weather working ability and strong penetrating power.SAR image change detection is a process of obtaining ground feature change information by qualitatively and quantitatively analyzing SAR images in the same area at different time phases.The SAR image has a large amount of coherent speckle noise attached to the image due to its own imaging mechanism,which brings great challenges to SAR image change detection.This thesis first introduces the current research status of SAR image change detection.Based on the analysis of the advantages and disadvantages of existing change detection methods,this thesis proposes a SAR image change detection method based on semantic convolutional neural network and generative adversarial learning,which mainly includes the following three research.Aiming at the problem of computational resource consumption caused by deep neural networks as the network layer deepens and the semantic details of images are easily lost in the downsampling process of semantic segmentation,this thesis proposes a lightweight CNN change detection method based on cascaded dilated convolution and attention mechanism.This method reduces the computational cost by the parallel initial module and the bottleneck structure of the encoder-decoder network,so that the amount of network parameters and floating points of operations are greatly reduced;through the proposed cascaded dilated convolution layer,the receptive field is increased while retaining the image semantic detail information,and compared with ordinary convolution,it will not bring additional computational cost.Putting it in encoder network and attention mechanism makes the image features more robust.This method has been verified by experiments on four datasets of real SAR images,and all of them have achieved good results.Aiming at the problem that the traditional change detection framework is sensitive to the quality of the SAR difference image,this thesis proposes a change detection method based on the siamese pyramid pooling network.This method uses the siamese neural network architecture to directly obtain the change information from the dual-temporal SAR image,thereby avoiding the influence of a large amount of noise in the difference image on the change detection result.The pyramid pooling module is added to the siamese feature extraction network to aggregate semantic contextual information of different scales,thereby improving the network's ability to capture global information.The dual-temporal feature maps extracted by the siamese neural network are added to form a fusion feature map,and the column vector is used as the input of the fully connected layer classification network,the final change detection result is obtained through the fully connected layer classification network.This method has been verified by experiments on four datasets of real SAR images,and all of them have achieved good results.Aiming at the problem that the noise in the SAR image difference image is difficult to suppress,this thesis proposes a change detection method based on the optimized difference image by the Generative Adversarial Network.This method uses the powerful style transfer ability of the Generative Adversarial Network(GAN)and the reference images marked in the training datasets to construct two GAN difference image optimizers to convert the difference image style into the reference image style to achieve the optimization of the difference image.The optimized difference image only needs a simple unsupervised classifier to get the final change detection result.This method is verified by experiments on four datasets of real SAR images,all of which can remove the noise of the difference image and retain the information of the change area in the difference image,so as to achieve the goal of optimizing the difference image and obtain good detection results.