Research On SAR Image Target Recognition Under Multi-view Observation

Synthetic Aperture Radar(SAR)is a microwave radar with the ability to penetrate clouds and fog.SAR is an active working mode sensor,which can realize all-day and all-weather monitoring of ground targets by actively transmitting electromagnetic pulses and receiving target echo imaging.Therefore,SAR is widely used in military,agriculture,surveying and mapping,disaster monitoring,and other fields.Multi-view SAR observes the target from multiple angles.Compared with the single view SAR observation mode,it can fully reflect the scattering characteristics of the target.With the increase of SAR equipment and the improvement of SAR image resolution,multi-view SAR image target recognition has become a new research hotspot.Based on the multi-view SAR image data and the characteristics of the multi-view SAR image,this thesis research the multi-view SAR image target recognition method,aiming to improve the recognition performance of the multi-view SAR image method in real situations.The research considers the factors that affect the performance of multiview SAR image target recognition,including the number of platforms,trajectory planning,test cost,imaging time,etc.resulting in sparse multi-angle SAR image data;it is difficult to accurately interpret the image target caused by the occlusion of the target during observation;The target image is difficult to mark without prior knowledge.The main works of this thesis are as follows:(1)The sample is limited under the sparse observation of multi-view SAR,and the target recognition method based on neural network is prone to overfitting.In this regard,the intra-class and inter-class relationship representation of multi-view SAR images are introduced in the feature space,and a loss function named Limited Data Loss Function(LDLF)is proposed.In addition,a feature combination plug-and-play module is proposed to make the LDLF loss function can be used in the current mainstream singlebranch network.(2)In complex scenes,the occlusion of the target leads to inaccurate target recognition results,and the applicability of the model decreases.In this regard,combined with the correlation between the adjacent azimuth angles of multi-view SAR images,a Tucker decomposition tensor completion algorithm based on multi-way delay embedding is proposed to learn the delay invariant structure and latent relationship expression between images.This algorithm can complete the missing information caused by occlusion,and effectively improve the performance of multi-view SAR image target recognition under occlusion.(3)The accuracy of data labeling will greatly affect the performance of the target recognition algorithm under the supervised learning framework.Unsupervised learning can adaptively mine data relationships without label information,which can effectively solve the problem of data labeling.In this regard,in view of the shortcomings of the Kmeans clustering algorithm that the number of clusters needs to be artificially set,combined with the similarity between multi-view SAR images,a method for estimating the number of clusters is proposed.It can effectively estimate the number of categories of unlabeled data sets.