Foliage-concealed Target Change Detection For UWB SAR

Ultra Wide-band Synthetic Aperture Radar (UWB SAR) can penetrate the foliageand image the concealed targets in it, which is very valuable for military applications.However, due to strong trunk clutter, foliage-concealed targets detection with UWBSAR images has been troubled with high false alarm rate problem for a long period. Forthis reason, the technique of foliage-concealed targets change detection based onMulti-temporal UWB SAR images is systematically and deeply researched in this thesis.The object of the research is to effectively suppress trunk clutter and improvefoliage-concealed detection performance by make use of strong correlation betweentrunk clutter in different temporal UWB SAR images. The main research works in thisthesis are listed as following:The problems existing in image preprocessing of UWB SAR foliaged-concealedtarget change detection are investigated. An image registration algorithm based on pointfeature is proposed. In the algorithm, taking into account of the characteristics of highnoise in UWB SAR images, a new approach of point feature extraction is designedbased on multiscle Harris operator, which improves the reliability of image registrationoperation under high nosie situation. Then the performance of SAR image relativeradiometric normalization approach based on linear regression model is theoreticallyanalyzed, and the error existing in it is pointed out as well. After that, a new relativeradiometric normalization approach is proposed. The approach is based on bi-directionlinear regression model, and the accuracy of SAR image radiometric normalization canbe improved by it.The problems existing in UWB SAR foliage-concealed target pixel level changedetection are investigated. Firstly, the change detection performance of image differencemethod and image ratio method are analytically compared under ideal conditions, andthe image difference method is proved to be superior to the image ratio method. Afterthat, an improved image difference change detection method based on imagesegmentation is proposed. Comparing with original difference change detection method,difference change detection method based on image segmentation can overcome theadverse influence brought by fast-fluctuating character of backscatter intensity infoliage area and correlation among multi-temproal images. Therefore, it has a betterperformance in practical application. At last, the difference change detection methodbased on image segmentation is further optimized to accelerate the speed of imagesegmentation and improve the accuracy of clutter distribution estiomation.The problems existing in UWB SAR foliage-concealed change detection based onstatistical distribution feature are investigated. A new statistical distribution featurechange detection algorithm based on generalized Laguerre polynomial is proposed. Comparing with existing algorithm, the algorithm based on generalized Laguerrepolynomial has a higher estimation accuray of probability density function. Therefore, ithas a better change detection performance. Besides that, a new statistical distributionfeature change detection algorithm based on bi-dimensional Edgeworth expansion isalso proposed. In existing algorithm, only one-dimensional probability density functionscomparision is implementd. However, in the new algorithm based on bi-dimensionalEdgeworth expansion, the two-dimensional probability density functions comparison isalso included. So a better change detection performance can be achieved by using thenew algorithm.The problems existing in UWB SAR foliage-concealed fusion change detection areinvestigated. In order to overcome the shortages of fusion change detectiom methodbased on Support Vector Data Description (SVDD), the Laplacian Eigenmap (LE)algorithm, which belongs to manifold learning field, is applied to improve SVDD, and aLaplacian Eigenmap SVDD (LE-SVDD) fusion change detection method is proposedbased on it. Comparing with the fusion change detection method based on SVDD, theproposed method not only extracts discriminant information from training set, but alsooptimizes the discrimination function according to the geometric structurecharacteristics of samples. Therefore, it has a better change detection performance.However, in practical application, LE-SVDD training process usually needs to solve theinversion of a large-scale martrix, which need a large amount of computation. In orderto deal with this problem, a LE-SVDD training method based on Nystr mapproximation is proposed, which effectively cuts down the computation burden andimproves the speed of fusion change detection.