SAR Images Change Detection Based On Generative Adversarial Network And Non-local Neural Network

Change detection of remote sensing images is a process of acquiring change information of ground objects by analyzing the differences of pixels,textures and structures between remote sensing images taken at different times but at the same place.Synthetic Aperture Radar(SAR)image has become an important source of remote sensing data because of its high resolution and powerful penetration ability.SAR images change detection is widely used in many fields such as environmental detection,land cover,urban planning,military strategy and so on.SAR images change detection is studied deeply in this thesis and the following work is accomplished around how to improve the accuracy of change detection:(1)A feature difference bi-discriminator adversarial network constrained by distance is proposed and applied to SAR images change detection.The network is made up of two discriminators.The first discriminator is used to infer the labels of the two time-phase samples,and the second discriminator is used to discriminate the authenticity of the labels.Two discriminators are adversarial thus promoting each other and improving each other.After training,the labels inferred from the first discriminator are so similar to the real labels that the second discriminator can not distinguish whether the labels are real or fake.In addition,considering the essential task of change detection is to find the differences between two time-phase image samples,the two time-phase training samples are input into the two branches of the first discriminator to extract their features respectively,and the differences between the two time-phase samples are calculated at the same time.The network learns the feature of changed areas,thus greatly reducing the false negatives.Finally,distance constraints are applied to the features between two time-phases image samples.By aggregating unchanged samples and separating the changed samples as far as possible,the network can learn more discriminative feature differences,and the labels are more accurate.Experiments on several real SAR image data sets demonstrate the effectiveness of this method.(2)A change detection method based on mutual non-local neural network is proposed in this thesis.Existing non-local neural network only computes the weighted average value of pixels in an image to capture the long-range dependences.Considering that change detection data often consist of two images,only establishing a non-local connection on a single image will not reflect the fundamental attributes of change detection.Therefore,a mutual nonlocal operator is proposed to calculate the mutual non-local weighted average value of two time-phase image samples,so as to capture the mutual long-range dependences between two time-phase samples,which can effectively improve the recognition ability of the network and reduce the influence of speckle noise in the original images.In addition,the mutual non-local operator is encapsulated into non-local block,and is used as the input layer of convolution neural networks which is named as mutual non-local neural network.Finally,the mutual non-local network is used for SAR images change detection.The experimental results prove that the mutual non-local neural network has excellent performance in images change detection.