Research On Generative Adversarial Technology Of Video SAR Images

Synthetic Aperture Radar(SAR)is less restricted by light and weather and can provide highresolution target images around the clock.Video SAR can continuously capture SAR targets at a certain frame rate,thereby presenting the dynamic information of the observed scene in the form of video,providing the possibility for real-time perception of the scene.However,since the video SAR technology has just started,the video SAR images that can be publicly obtained are very limited,which brings great difficulties to the research of related scene perception technology.In view of the current situation,this paper adopts the generative adversarial technology to study the generation method of video SAR images.The main research results are as follows:(1)Studying the decoupling control mechanism of the image generators.Decoupling control can map different elements of the generator input vector to different image features,ensuring that the elements of the control vector are independent of each other,so that each feature of the generated image is controlled by a specific element,and then implement feature decoupling for generated images.(2)Building a Video SAR Generative Adversarial Network.This network draws on the improvement methods commonly used in generative adversarial networks,and adds a mapping network to the generator that can decouple the control vector to realize the control and adjustment of the generated image features;the generative network adopts the method of step-by-step generation and adds a style module,so that the generated SAR image has richer detailed information.The dataset required for network training is obtained by object extraction from real SAR entity images and shadow images using image labeling scripts.The experimental results show that by adjusting the value of a certain element in the control vector,the corresponding features of the SAR image can be changed,the video SAR generative adversarial network has a good feature decoupling effect,and the final generated SAR image sequence is closer to the real image sequence.(3)Building a video SAR image database.The spherical linear interpolation is performed on the real image and the shadow image to obtain the SAR image sequence.The target center of each image in the sequence is filled corresponding to the motion trajectory,so that the generated image has the same resolution as the real SAR background image,and then image fusion is achieved by pixel replacement.The trajectory of the target in the experiment is randomly generated and affected by Gaussian disturbance and deterministic inertial components.In order to make the generated image sequence closer to the real image captured by the radar,the fused image is rotated and cropped to ensure that the background rotates at a uniform speed in the fused image sequence,and the target also moves relative to the background.Through batch fusion,a large number of video SAR image sequences with ground moving targets in different backgrounds are obtained,and a video SAR image database is established to provide data support for subsequent target tracking and detection.