Several Classes Of Multiscale Nonlinear Random Models And SAR Image Bootstrap Segmentation

With the appearance of SAR (synthetic aperture radar) platform for multiple-pattern, multi-band, full polarimetric, 3-D imaging and moving target detection, and development multiple frequency, multi-polarization and multi-vision technique. SAP, image segmentation is a key technique of automatic target recognition (ATR) and information processing. The existence of speckle in SAR imagery makes segmentation more difficult than that of optical image. We present several new segmentation methods, including Bootstrap spatially variant mixture multiscale autoregressive prediction (SVMMARP) model for image segmentation, two improving multiscale likelihood ratio method for supervised and unsupervised image segmentation, Bootstrap generalized multiresolution likelihood ratio (GMLR) for supervised SAR image segmentation, three unsupervised image segmentation methods based on GMLR, and spatially variant mixtures of multiscale autoregressive moving average (SVMMARMA) model for image unsupervised segmentation. The contribution of the the thesis is as follows:Firstly, we present SVMMARP model and its improving algorithm is proposed. The SVMMARP model is not only capable of describing spatially variant characteristics but also exploits multiscale autoregressive statistical properties of SAR imagery and can reduce the effect of speckle. The performance of the SVMMARP model has precise segmented results in comparison with classical mixture model and MAR model; and a kind of method selecting component number quickly is proposed. Improving SVMMARP model method considers both CPU time and precision of parameters. Experimental results shows that improving SVMMARP model method has smoother edges, isn't sensitive to speckle noise and gets more pricise segmentation in comparison with classical mixture model and MAR model.Secondly, a new model, SVMMARMA model, is present. We derive its EM algorithm for parameter estimation, and get the satisfying results when this new model is used for SAR image segmentation.Thirdly, we present the definition of generalized multiresolution likelihood ratio, which has the advantage of classifying different kinds of signals more accurately than classical likelihood ratio by fusing more and different signal features. We also give a supervised SAR image segmentation method, and get segmentation results by the GMLR histogram based on a judicious window size. We propose three unsupervised SAR image segmentation methods based GMLR. The three methods consider the Markov or relativity of neighborhood pixels. Performance results show the threemethods are not sensitive to speckle noise and get more pricise segmentation incomparison with relative models.Fourthly, a new multiscale expression is derived, and we give a supervised SARimage segmentation method and an unsupervised image segmentation method basedon hypothesis and test theory; from improving multiscale expression, improvingmultiscale expression fuses more features than previous multiscale expression. Experimental results shows improving multiscale expression has more pricise segmentation.