Application of unsupervised segmentation for SAR imageries based on multiscale stochastic models
Yi Xiong, Jin Jin Ding, Wei Wang · 2016
A new unsupervised segmentation algorithm of SAR(Synthetic aperture radar) imageries based on multiscale Stochastic Models is proposed, considering non-gaussian statistical property of SAR image data and Markov property of neighboring scales. Since EM(expectation maximum) algorithm can not get the parameter estimation of non-gauss distribution, MAR(Multiscale Autoregressive) model is used for extracting image Feature data which obeys gauss distribution. HMT(Hidden Markov Tree) model can be used to model image consisting of multi-scale feature data, which can be approximated by mixed gauss distribution and its parameters can be straightly trained by EM algorithm. Then we propose a context model to fuse feature information of multiscale. Finally, we obtain a new unsupervised segmentation approach for SAR imageries. Simulations on SAR imagery indicate that the new approach improves segmentation accuracy in some degree.