Saliency Detection Based On Regional Statistical Characteristics In SAR Image

Synthetic Aperture Radar(SAR)is a kind of active microwave imaging remote sensing radar with obvious advantages,which can flexibly adapt to weather changes,penetrate obscured obstacles,and achieve all-day,all-weather imaging.Therefore,SAR is widely used in civil and military earth observation missions,and the research of detection technology for SAR images has attracted much attention.Visual saliency-based detection techniques can quickly narrow down the candidate region of the target to be detected and help improve the accuracy of the target detection,which is one of the popular research topics in the current detection field.However,traditional saliency detection methods still have certain limitations in the applications of SAR images.On the one hand,the commonly used saliency features lack in-depth consideration of the imaging mechanism and statistical characteristics of the target in SAR images.Most of these features are designed for optical images,and the chosen features such as color and luminance contrast cannot fully exploit the available information of SAR images.On the other hand,there is serious coherent speckle noise in highresolution SAR images,which interferes with the original data features.Detection using pixel-level features cannot effectively resist such interference,and often suffers from reduced detection accuracy and reduced efficiency.After analyzing the above problems,this paper focuses on two different polarization methods of SAR images with single polarization and multiple/dual polarization.From the perspective of statistics,we dig deeper into the saliency features based on the scattering characteristics of SAR images and study the saliency detection method based on the statistical characteristics of the target at the regional level.The main content and innovation points are as follows.(1)Aiming at the problems of in the saliency detection in single-polarized SAR images,such as the problems that the commonly used saliency features don’t make full use of the differences in scattering statistical properties between the target and background and the low efficiency and poor noise immunity of the pixel-level algorithm,we propose a saliency detection method based on the superpixel-level lacunarity feature for single-polarized SAR images.According to the backscattering mechanism of man-made targets and natural terrain background to study the differences between their statistical characteristics,we propose the saliency feature based on the lacunarity,design two channels inside and outside the superpixel to comprehensively evaluate the saliency of the image,and use the prior information of the target to post-process the obtained saliency map to further suppress the background region.The method realizes the saliency detection for vehicle targets in SAR,which effectively improves the detection accuracy.At the same time,the introduction of superpixel units enhances the noise immunity of the algorithm.Compared with other methods,the average precision of this method is improved by about 0.3.(2)Aiming at the problems in the saliency detection of polarized SAR images,such as insufficient utilization of polarization information,poor regional consistency of detection results and serious interference from noise in pixel-level detection,we propose a saliency detection method based on the superpixel-level likelihood ratio test for polarized SAR images.By analyzing the stability difference of scattering characteristics between man-made targets and natural features over time to design saliency features based on the statistical distribution characteristics of polarized SAR data.We firstly construct the statistical distribution model of dual-polarized SAR data at superpixel level,and then design the saliency feature and its detection method based on the superpixel level likelihood ratio test.The method achieves effective detection of the builtup area of polarized SAR images,improves the accuracy of detection.Compared with other methods,the OA is improved by about 0.2.And it obtains detection results with stronger integrity and clearer boundary of the target area.