Research On Target Characteristics Modeling And Detection Method Of High-Resolution SAR Image

Target detection of Synthetic Aperture Radar(SAR)images is used to discover regions of interest in imaging scene.Results of detection provide important information for both military reconnaissance and civil monitoring,and they are also regarded as inputs for target classification and recognition.Otherwise,the poor detection will seriously degrade the performance of these subsequent procedures.Hence,target detection is an important part of SAR image interpretation.High-resolution(HR)and multi-polarization SAR images provide abundant information of targets.However,some complex characteristics are observed in the target regions in HR SAR image,such as heavily fluctuant amplitudes,disconnected strong scattering,and mixed polarimetric components.Most of the traditional detection methods only exploit the characteristics in the background,whereas the characteristics of the targets are ignored.Consequently,they might not extract adequate information of the targets.In order to improve the detection performance of HR SAR images,the modeling of the targets' characteristics is carried out in this thesis.Specifically,the characteristics in this thesis refer to scattering characteristics,geometrical characteristics and polarimetric characteristics.For these characteristics,their influence on target detection is analyzed,and several detection methods are proposed.The main contributions are as follows:1.Common characteristic models of targets and background in SAR image and their differences are analyzed,and some detection methods are introduced.Then,CFAR detection results from SAR images of different resolutions are analyzed based on template matching,and SAR images with different polarizations are also evaluated based on polarimetric decomposition theory.2.To deal with fluctuant amplitudes in each target,a conditional entropy is proposed to describe the scattering characteristics,and a detection method is built based on iterative outlier detection and recursive saliency optimization.The proportion of strong and weak scattering in the target is regarded as the parameter of the conditional entropy,that maximizes the Likelihood Ratio Detection(LRT)for the target.Then,the iterative detection and recursive optimization can modify saliency of outliers,and solve the problem of background estimation.Hence,the proposed method can improve the performance of target detection in complex SAR images.3.To combine separated regions in each target,a part-based model is proposed to exploit the geometrical characteristics,and a target detection method is proposed by searching parts.The detection with part model consists of three cascades: strong scattering regions are extracted as nodes of interest;the nodes are selectively searched based on their geometrical correlations and similarities,and used to generate parts;the parts are greedily combined based on geometric constraints to form potential targets.By using the geometrical characteristics,isolated nodes and parts are excluded as false alarms.4.For the mixed polarimetric components in each target,a classification model of the polarimetric characteristics is built,and a saliency detection method is built based on globally weighted perturbation filters.The classification of polarimetric components is useful for simplifying and improving target detection.Meanwhile,if targets are assumed to be sparse,saliency can be computed for the classes.Then,saliency detection is achieved by fusing the responses from the weighted filters.Saliency analysis about the mixed polarimetric components is useful for overcoming their negative impact on target detection and depressing false alarms that are caused by the difference between polarimetric characteristics of some background classes.This thesis extended target detection methods by analyzing the target characteristics in HR SAR images,and experiments based on HR SAR images verified the performance of the proposed methods.