Single Loop Inference of Hidden Markov Tree for Multiscale Image Segmentation

This paper addresses the problem of bottom-up and up-bottom multiscale segmentation of objects in the presence of dynamic backgrounds. Previous hidden Markov tree (HMT) based approaches have exploited an iterative inference scheme and each iteration consists of an two-stage segmentation mechanism, namely, parameter learning and multiscale fusion of likelihoods. In this paper, we propose a novel approach for recovering multiscale segmentation accurately in the absence of iterative multiscale fusion stage. This allows both inference and fusion of multiscale classification likelihoods to be computed in a single loop through bottom-up likelihood estimation and up-bottom posterior inference of HMT. Experimental results on a synthesized image in the presence of Gaussian white noise demonstrate the high robustness achieved by the proposed method.