Research On Automatic Target Extraction In SAR Imagery

Owing to its all-day and all-weather work characteristics,synthetic aperture radar(SAR)has become an indispensable remote sensing device,and plays an important role in both of the military and the civilian fields.The automatic target extraction in SAR imagery is one important application of the SAR instruments.And the research on this technology has been widely concerned by scholars both at home and abroad recently.With the purpose of developing practical technologies for SAR image automatic target extraction,this dissertation makes a systematic and in-depth study on several key problems in this field based on its relevant principles and application backgrounds.These key problems include target detection in complex scene SAR images,target discrimination in complex scene SAR images,and target chip segmentation used for target discrimination features extraction.The main research efforts can be summarized as the following aspects:1 The first part focuses on the problem of target detection in SAR images under complex scenes with multiple targets and clutter boundaries.For this tough problem,the selection of the homogeneous pixels in the background reference region is the key factor in the problem solving,which is usually implemented with pixel censoring or half-window censoring strategies.By combining the respective advantages of the traditional pixel and half-window censoring detection algorithms,a novel constant false alarm rate(CFAR)detection algorithm based on automatic block-to-block censoring is proposed under the assumption that background clutter follows theG~0 distribution model.At the beginning of this detection method,the local reference detection window is first partitioned into several blocks with the same size.And then the variability index statistic is used to censor the blocks in the local reference window in order to reject the nonhomogeneous ones in which there exists nonhomogeneous pixels.Then the mean ratio statistic is utilized to select and merge the homogeneous blocks which have the same distribution,in order to solve background clutter censoring problem in clutter edge situations.At last,with the selected reference clutter block regions,the parameters of the background clutter statistical model are estimated,and then the binary detection is implemented for the pixels in the block under test.2 In the second part,from the viewpoint of image superpixel segmentation,the problem of target detection in complex scene SAR images is further studied.With gradually increasing of the resolution of SAR images,the local structures of targets in SAR images have become more and more obvious.Superpixel algorithm can be used to group local pixels with similar properties in the image.And so the local structures of the targets in SAR images can be properly separated from the background clutter.To use superpixels instead of pixels as the unit of an image,not only can improve the computational efficiency of the subsequent detection methods,but also can improve its final detection performance.The work mainly has the following two points:(1)Under the assumption that the local region clutter follows the Gamma distribution,a SAR-SEEDS superpixel segmentation algorithm is proposed.The algorithm starts from a regular grid partition of the image as the initial superpixels.And then it iteratively refines the superpixels by modifying the boundaries in an image hierarchy according to the decision criterion based on an energy evaluation function of image superpixel segmentation.The SAR-SEEDS method first performs edge refinement in a large-scale layer to achieve coarse-edge modification of the superpixels,and then gradually reduces the scale.Finally,edge refinement is performed in the pixel layer to achieve fine-edge modification of the superpixels.(2)On the foundation of superpixel segmentation preprocessing of SAR images and under the assumption that the background clutter follows theG~0 distribution model,a target detection algorithm based on superpixels censoring and merging is proposed.The algorithm implements target detection through the following steps:Firstly,the SAR-SEEDS method is used for superpixel segmentation of the SAR image under test,in order to realize the local uniform area partition and separate the targets from clutter regions;Secondly,with a superpixel censoring algorithm,the superpixels of the image are separated into two classes:background clutters and potential targets;Thirdly,in the reference region of a superpixel under test,the background class superpixels are merged using an fast superpixel merging technique to achieve image segmentation of the background area;Next,according to the image segmentation result,the reference background clutter is selected properly,and the center superpixel detection is subsequently performed.At last,the target-level detection result is obtained with another superpixel merging method.3 The third part focuses on the problem of target discrimination in complex scene SAR images.Considering that the distributions of the local structural features in the target and clutter chips are different,a SAR target discrimination algorithm based on the multi-feature fusion bag-of-words(BOW)model is proposed.In the low-level feature extraction stage of the BOW model,the SAR-SIFT feature is utilized to describe the shape information of local patches of an image sample.And also,a set of new local descriptors is used to capture the contrast information and the texture information of the local patches,which is extracted based on the traditional target discrimination features.For the fusion of different low-level features in BOW model,the image-level feature fusion strategy is implemented to generate the image global feature,which is realized by a multiple kernel learning(MKL)method with L2-norm regularization.4 The forth part focuses on the problem of image segmentation for SAR target chips.For this problem,the coherence of pixels amplitude in the background region and the target region has a direct influence on the image segmentation result.Considering this,an Otsu segmentation algorithm for SAR images based on power transformation is proposed.In the proposed algorithm,firstly a speckle suppression method for SAR images is applied to improve the coherence of the background region and the target region.Secondly adaptive power transformation is employed to the filtered SAR image to enhance the coherence of the target region furthermore.At last,the original one-dimensional Otsu method is adopted to segment the transformed SAR image.