Aircraft Target Detection In Very High Resolution SAR Images Based On Information Interaction And Migration Learning

Synthetic Aperture Radar(SAR)image interpretation has been widely concerned by researchers.Especially in recent years,with the development of radar technology,the resolution of SAR image is getting higher and higher,and even reaches very high resolution,which brings opportunities and challenges to interpretation.In the very high resolution SAR images(Mini SAR),target surfaces and lines often present the following phenomena: different sizes,different scattering intensities and disconnected bright spots make the target incomplete,which makes direct target detection extremely difficult,and leads to the problem of low detection rate.Aiming at the above problems,this paper presents a method of aircraft target detection in very high resolution SAR images based on information interaction and migration learning.The main works are as follows:(1)A method of acquiring candidate regions of interest targets based on bright-dark threshold segmentation and k-nearest neighbor selection is proposed.Firstly,according to the characteristics of the aircraft target in the SAR image,the SAR image is separately subjected to bright threshold segmentation and dark threshold segmentation to obtain bright region candidate map and dark region candidate map.Then,according to the prior knowledge that a target shadow may correspond to multiple bright spots,k bright regions nearest to dark regions are selected and merged with dark regions to obtain candidate regions of interest in pixel space.Experimental results show that this method can effectively extract candidate regions of interest in very high resolution SAR images.(2)A method of aircraft target shadow completion based on wing shadow structure is proposed.Firstly,candidate regions of interest in the pixel space are corresponded to the semantic sketch map of SAR image to obtain candidate regions of interest in the semantic space;secondly,sketch segments of suspected wing shadows in candidate regions of interest in the semantic space are extracted based on the neighborhood structure relationship of sketch segments;finally,analyze the wing shadow structure in the area of the aircraft target candidate regions containing the suspected wing shadow sketch segments,using different shadow completion methods for different regions,one is the method of shadow completion for the region where both sides of the wing shadow are complete,and the other is the method of shadow completion for the region where only one side of the wing shadow appears.The application scenarios of the two methods are different,but they can all complement the shadow of the aircraft target region and get the target shadow region map.This method combines the gray level information of the target in the pixel space and the structure information in the semantic space.The experimental results show that the candidate regions of the aircraft target can be obtained and the shadow of the target can be filled by the interaction of the two spatial information,which is helpful for the classification and recognition of the candidate regions of the aircraft target in the later stage.(3)A method of classification of aircraft target candidate regions in very high resolution SAR images based on migration learning is studied.Firstly,the binary template of different aircraft targets in natural images is designed and used as the source domain.The target shadow regions of the target to be classified after shadow completion are processed,and the binary images of the target shadow regions are obtained,which are regarded as the target domain.By the Generative Adversarial Network migration learning method based on sketch labeling information,the label information of the source domain is assigned to the target domain data.The task of classification of candidate regions aircraft target is accomplished.The migration learning method can solve the problem that SAR image data is small and labels are difficult to obtain,which can realize the accurate classification and recognition of aircraft target candidate regions in SAR image.