Bootstrapping Stochastic Annealing EM Algorithm for Multiscale Segmentation of SAR Imagery

Abstract

A valid Bootstrapping stochastic annealing Expectation Maximization (BSAEM) algorithm is proposed for unsupervised and multiscale segmentation of synthetic aperture radar (SAR) imagery. Given an original SAR images, we construct multiscale sequence of SAR imagery and randomly select a small representative set of pixels based on Bootstrap technique, which reduces the dependence effect of pixels in SAR imagery. Then, mixture multiscale autoregressive model (MMAR) is employed for modeling SAR imagery, and BSAEM algorithm is proposed for SAR imagery segmentation. The algorithm consists of four steps, namely Expectation Step, Stochastic Step, Annealing Step and Maximization Step, which not only modifies convergence property of classical Expectation Maximization (EM) algorithm but also reduces the segmentation time. Finally, some experiment results are given based on our proposed approach, and compared to that of the EM algorithm. The results show that our algorithm gives better results than the EM algorithm both in the quality of the segmented image and the computational time.