SAR Image Despeckling And Segmentation Based On Statistical Model

Synthetic aperture radar (SAR) image despeckling and segmentation, which areessential to the exploration of SAR image, are significant steps towards SAR imageinterpretation. They establish the basis of automatic target recognition and promoteSAR applications in many fields. During the last decades, the two works have beenwidely studied.This dissertation studies the SAR image despeckling and segmentation based on theresearch of particle filter (PF), fuzzy clustering, multiresolution analysis, Bayesianfusion theory and triplet Markov fields model, given the satellite-borne and airborneSAR images.The main contents of this dissertation are summarized as follows:The first section proposes a modified particle filter (PF) despeckling algorithm instationary wavelet transform (SWT) domain based on the study of the statistics of SARimages. PF algorithm is an effective approach to nonlinear and non-Gaussian Bayesianstate estimation and it has been successfully applied to SAR image despeckling. First,we show that the wavelet coefficients of SAR images which exhibit significantlynon-Gaussian statistics can be described accurately by generalized Gaussian distribution(GGD) in stationary wavelet domain. Secondly, to amend the weight deviation, Markovrandom field (MRF) is introduced to redefine the weight of the particles. Furthermore,the sampling interval is updated according to the new weight. At last, region-dividedprocessing is implemented for the real time application of the proposed algorithm.Experimental results of the simulated images and real SAR images despecklingdemonstrate the ascendant performance of the proposed algorithm in noise reduction,preservation of the edges, single target and textural features of SAR images.In the second section, based on the studies of the fuzzy clustering theory andnonstationary statistical property of SAR images, we propose a statistical fuzzy tripletMarkov fields (FTMF) model, which is a fuzzy clustering type treatment of tripletMarkov fields model, for unsupervised multi-class segmentation of SAR images. Thissection contributes to SAR image segmentation in four aspects:1) Mean fieldapproximation (MFA) theory is generalized to deal with planar variables to derive priorprobability in FTMF model, which resolves the problem in Gibbs sampler in terms ofcomputation cost.2) Nonstationarity of the statistical distribution of SAR intensity/amplitude data is taken into account to improve the spatial modeling capabilityof FTMF model.3) A fuzzy objective function of FTMF model is constructed for SARimage segmentation.4) Parameters estimation methods with respect to FTMF modelparameters, fuzziness and statistical distribution parameters of SAR data are proposed.Experimental results and analysis of simulated data and SAR images demonstrate thatthe proposed algorithm is able to retain more information from SAR image, thusimproving the edge location and region homogeneity effectively.The third section proposes a multiscale and multidirection triplet Markov fieldsmodel in wavelet domain, named as WTMF model, based on the studies of themultiscale statistical modeling and statistics of SAR images. In WTMF model, amultiscale causal WTMF energy function is constructed to capture the intra-scale andinter-scale dependencies in label and auxiliary fields. And multiscale likelihoods ofWTMF model are derived based on wavelet hidden Markov tree (WHMT) to capturethe statistical properties of wavelet coefficients. The proposed model can integrate theglobal and local information in terms of spatial configuration and image features in amore complete manner. The coarser-scale information is utilized to guide the finer-scalesegmentation. Moreover, we analyze the nonstationarity of SAR images from thetextural point of view in a multiscale framework and present a reasonable initializationof auxiliary field based on the nonstationary anisotropy Gaussian kernel (NAGK)parameters. Experimental results and analysis have proven that the global structuralinformation at coarser scales guides the segmentation in fine scale; thus WTMF modelcan resolve the mis-segmentations effectively and is more robust against speckle.In the forth section, based on the studies of the hierarchical statistical modeling andstatistics of SAR images, we propose a hierarchical triplet Markov fields (HTMF)model for unsupervised SAR image multiclass segmentation. In HTMF model, theintra-scale clique potential and the conditional clique potential are utilized to captureintra-scale interactions and the inter-scale dependencies within label and auxiliary fields.In virtue of Bayesian inference on the quad-tree, HTMF model captures the global andlocal image characteristics more precisely in the bottom-up and top-down probabilitycomputation. In this way the underlying spatial structure information is propagatedeffectively. In the frame of experiments considered, HTMF model works better than theclassical TMF model. HTMF model is more robust against speckle and achieves bettersegmentation results for nonstationary SAR images while preserving the spatialstructure. The fifth section proposes a conditional triplet Markov fields (CTMF) model forunsupervised multiclass SAR image segmentation based on the studies of thenonstationary property, statistical modeling and textural feature of SAR images. CTMFmodel considers the nonstationary property of image from the perspective of spatialconfiguration and statistical distribution of SAR intensity/amplitude data explicitly.Normalized logistic regression model (NLRM) is utilized to construct associationpotential function, and thus consider the dependence of the observed data. CTMF modelconstructs the data-dependent interaction potential function in label fields to capture thespatial correlation more precisely. The experimental results of simulated data and realSAR images segmentation demonstrate that CTMF model is suitable for dealing withunsupervised segmentation of SAR images with complex textures.