Study On Man-made Target Detection Method For SAR Imagery In Complex Scenes

Synthetic aperture radar(SAR)has high military and civil value because of its high resolution,all-day and all-weather characteristics.Ship target and building target,as typical sea maneuvering target and land fixed large range target respectively,their detection methods have been widely concerned by scholars.As a state-of-the-art method,the constant false alarm rate(CFAR)detection algorithm does not need a large number of training samples,and is of great significance for ship detection in some specific scenes.However,in high resolution SAR images,the detection performance of the classical CFAR is easily affected by speckle noise.The detection results are sensitive to the size of the sliding window.It is difficult to ensure that there are no target pixels in the cluttered background,and the computational efficiency is low.Extraction of target scattering information in land area is the key to SAR building detection.Because of its powerful feature analysis ability and high operating efficiency,deep learning has become a hot method for building detection in complex large scenes of SAR.Conventional building detection methods based on deep learning have less advantages in analyzing the scattering characteristics and electromagnetic characteristics of SAR targets.The fusion of deep feature and the generalization performance of the model need to be improved.This thesis studies the detection methods of man-made targets for SAR images in complex scenes.For ship targets,this thesis focuses on the CFAR ship detection method combined with the superpixel information of SAR images.For land building targets,this thesis focuses on the research of building detection method of large scene dual-polarized SAR combining artificial design features and depth features.The main research contents are as follows:(1)This thesis proposes a fast non-window CFAR ship target detection algorithm based on superpixels.The superpixel generation method of density-based spatial clustering of applications with noise(DBSCAN)is used to generate superpixels for SAR images.Under the assumption that SAR data obey the Rayleigh mixture distribution,this thesis define a superpixel dissimilarity measure.Then,the clutter parameters of each pixel are accurately estimated using superpixels,which can avoid the shortcomings of the traditional CFAR sliding window even in the case of multiple targets.A new local contrast based on the coefficient of variation(Co V)of the SAR image is proposed to optimize the CFAR detection result,which can highlight the difference between the ship target and the background,then eliminate a large number of false alarms.The experimental results on several SAR images show that this method has robust ship target detection performance.(2)This thesis proposes a pseudo-siamese dense convolutional pyramid network(PSDPNet)for building detection in complex large scenes of dualpolarized SAR.Scattering information and artificial design features are used as the input of dual-channel network respectively to extract different levels of features independently.To solve the problem of multiscale distribution of building area,a feature pyramid module is designed and the multi-level information fusion is carried out by combining the features extracted from the dual-channel network.Focal Loss is used to solve the problem of uneven distribution of positive and negative samples.Finally,the pixel-level dualpolarized SAR building area detection results are obtained.The experimental results show that the proposed method has efficient detection performance and good generalization performance.