A Trainable Hierarchical Hidden Markov Tree Model for color image segmentation and labeling

In this paper we consider how to annotate or label regions of grey-level or multispectral images based upon known labels and a set of interacting hierarchical doubly stochastic processes. The proposed model extends current work on the use of hidden Markov models for image processing using multiscale representations. In this paper we explore a new bijective up-down algorithm whereby the spatio-spectral context of specific image region signatures are encoded via different types of trainable support kernels for the upward and downward hidden Markov trees.