Local parameter estimation and unsupervised segmentation of SAR images

Our work deals with the unsupervised statistical segmentation of SAR images. However the method here developped is a general parameter estimation technique and can be used for most types of images. We adopt a contextual method in which each pixel is classified from the measurements taken in its neighborhood. In this approach the previous statistical problem is the estimation of components of a distribution mixture. We showed in some previous studies that the SEM is well adapted to the problem in this frame, when stationary random fields are considered. In this paper we present a new distribution mixture estimator in which priors can depend on the position of the considered pixel. This makes it valid in the non-stationary case. We describe some situations, based on synthetic images sampled by stationary or non stationary random fields, in which the contextual method based on parameters estimated by our algorithm is more efficient than the same method based on parameters estimated by the SEM algorithm.