Research On Shadow Detection Algorithm Of Video SAR Targetbased On Deep Learning

Video Synthetic Aperture Radar(VideoSAR)is a new imaging mode of Synthetic Aperture Radar(SAR).It not only has the advantages of all day and all-weather of traditional SAR,but also has the characteristics of high imaging frame rate and high resolution.It can realize the continuous and fast imaging of the monitoring sce ne and obtain the video imaging results,which can provide help for the real-time detection and tracking of targets in modern military.Since the research on VideoSAR target detection is still in its infancy,the traditional VideoSAR target detection methods have some problems,such as difficult inter frame registration,unclear shadow features of fast moving targets,complex detection steps,slow detection speed,and difficult to meet the real-time requirements,so this paper proposes a video synthetic ape rture radar target detection based on deep learning,which can achieve the same effect as the traditional methods,but the detection speed is greatly improved.The related work and progress are as follows:Firstly,the imaging principle of video SAR in circular mode is analyzed.At the same time,in order to better study the detection method of moving target,the imaging characteristics of moving target are studied.The Doppler frequency shift caused by the velocity of moving target and the shadow formation principle of static and moving target are analyzed.The position of target is determined according to the shadow position in different states.Based on the principle of moving target shadow formation,target level annotation and pixel level annotation are carried out for videosar data respectively.Secondly,a new target detection algorithm based on YO LOv3 is proposed,which applies the target detection algorithm in the image field to radar image.With the introduction of YOLO v3-tiny network structure,the amount of calculation is reduced while the accuracy is guaranteed,and the real-time requirement of detection is realized.The attention mechanism module is embedded to increase the proportion of vehicle target features,and experiments are carried out with manually annotated VideoSAR data sets.The experimental results show that the network with attention mechanism module is a network with balanced speed and precision.Finally,a method of vehicle shadow detection based on semantic segmentation is proposed.Three dimensional convolutional neural network model is used to extract the spatial and temporal features of continuous image sequence in video SAR,and the residual structure is used to reduce the network parameters and improve the learning ability of the network.The hole convolution is added to increase the receptive field,and the pixel level video SAR data is manually labeled.Experimental results show that the proposed method can detect vehicle targets in videosar accurately.