ADAPTIVE MIXTURE ESTIMATION AND UNSUPERVISED LOCAL BAYESIAN IMAGE SEGMENTATION

This paper addresses mixture estimation applied to unsupervised local Bayesian segmentation. The great efficiency of global Markovian-based model methods is well known, but the efficiency of local methods can b e competitive in some particular cases. The purpose of this paper is t o specify the behavior of different local methods in different situati ons. Algorithms which estimate distribution mixtures prior to segmenta tion, such as expectation maximization (EM), iterative conditional est imation (ICE), and stochastic expectation maximization (SEM), are stud ied. Adaptive versions of EM and ICE, valid for nonstationary class fi elds, are then proposed. After applying various combinations of estima tors and segmentations to noisy images, we compare the estimators' per formances according to different image and noise characteristics. Resu lts obtained attest to the suitability of adaptive versions of EM, ICE , and SEM. Furthermore, the local methods turn out to be robust in the sense that the parameter estimation step does not degrade the final s egmentation results significantly, and the choice of EM, ICE, or SEM h as little importance. (C) 1995 Academic Press, Inc.