Unsupervised Context-based Fused Segmentation Of SAR Image Based On Multiscale Markov Model

Due to the enormous quantity of data and the speckle, the segmentation methods used for optics images can not fit for SAR images. So efficient segmentation algorithms are very essential for recognition and interpretation of SAR images. This thesis describes some theories which are used for fixing the segmentation number and some approaches for the segmentation of SAR images in the framework of Multiscale Markov model.Firstly, we present a MMDL (Multiscale minimum description length) criterion which can be used for making sure the segmentation number of SAR images. This criterion is based on MDL which has been widely used in image segmentation. We have proved that it is efficient and correct to get the segmentation number of SAR images.Secondly, enlightened by the segmentation method in wavelet domain, we proposed a new unsupervised context-based fused segmentation algorithm based on Multiscale Markov model. This method fully considers the information on the Multiscale and approximates segmented imagery with mixture distribution. Now we have developed the context-based fused segmentation based on MGMM(Multiscale Gaussian Markov Model) or MRMM(Multiscale Rayleigh Markov Model).The model parameters can be straightly trained by T-EM(expectation maximum on the tree) or ICE(iterative conditional estimation) algorithm based on segmented image.Finally, we present a unsupervised segmentation algorithm of SAR images based on two Multiscale stochastic models. This method considers both the statistic characteristics of SAR images and the information on the multiscale, it solved the problem that EM algorithm can not be used to estimate the parameters of Multiscale Markov Model when approximate distributions are not the Gaussian distributions.Large numbers of experimental results show that the approaches presented in this thesis will be efficient and helpful for the segmentation of SAR images.