Research On The Evaluation Techniques Of Similarity Measure In SAR Images

Similarity measure, which measures the extent of similarity of two compared SAR images,is the basis of SAR image interpretation applications such as SAR simulated images evaluation,selection of SAR imaging data, target classification and recognition, image registration, changedetection, etc. However, by reason of the speckle noise in SAR images, target partition occlusion,fuzziness, image local gradient variation, the similarity measure evaluation becomes more andmore difficult. According to this problem, this thesis thoroughly studies the techniques ofsimilarity measure evaluation of SAR images based on gray-level and feature, furthermore, weanalyze the subjective similarity utilizing fuzzy mathematics. The main work includes thefollowing aspects.(1). Beginning with the pixel intensity of SAR images, the similarity based on SAR imagegray-level is defined. The characteristics of speckle noise and geometry are analyzed,furthermore, the uncertainty of image match is discussed and a new similarity measure based onpixel difference code is proposed. The code image is obtained by averaging the difference ofbrightness in adjacent pixels in SAR images. Then, the similarity measure for uncertain SARimages is obtained by evaluating the concordance of code images. The advantages of thissimilarity measure are obvious. Firstly, the proposed method is proved to be robust to specklenoise, partially occlusion and fuzzification. Secondly, using the proposed method, the SARimage matching result at confidence level is discussed.(2). The pixel intensity can not indicate the true scatter of target, especially in the regionwhere speckle noise is serious. Therefore, the similarity only based on pixel intensity is far fromsatisfactory. Hence, The similarity based on local texture feature is defined. We analyzed the twocommon pixel relativity measurement, including difference distance and ratio distance, therobustness to SAR speckle noise is discussed. With the analysis of speckle noise and localgradient variation in SAR images, a new similarity measure based on local gradient ratio patternis proposed. Firstly, we extract the gradient ratio pattern for each pixel based on Weber’s law.Secondly, the local gradient ratio pattern histogram (LGRPH) is computed. Finally, the similarityis obtained by utilizing K-L discrepancy to measure the distance of LGRPH. The proposedmethod is theoretically proved to be robust to speckle noise, and the adaptability to local gradientvariation is also discussed. Experimental results based on simulated and real SAR imagesdemonstrate that the proposed similarity measure is valid.(3). The aforementioned similarity measures which are based on pixel intensity or localtexture feature are all category of objective similarity. Here, The subjective similarity measure ofuncertain contours in SAR images based on fuzzy mathematics is defined. The common segmentation algorithms of SAR images are analyzed, besides, the existing problems areindicated. A confidence interval of similarity and the method of constructing credibility areproposed based on the characteristics of image contours. Firstly, the similarity measure isobtained by fuzzifying the uncertain contours in SAR images. Then, the confidence interval ofthe measure as well as the definition of credibility at a given confidence level are got byanalyzing the distribution function of fuzzy model. Using fuzzy mathematics, this methoddescribes the uncertainty of contours better. The proposed similarity measure between a givenpair of images is robust to the location of image contours. Besides, it can also provide reasonablesimilarity with credibility of the fractured and multi-edged contours, which has not beenpresented by other methods. Moreover, it is consistent with human visual perception.