Tree-structured belief networks as models of images

In this paper we deal with the use the tree-structured belief network (TSBN) as a prior model in segmenting a natural image into a number of predefined classes. The TSBN was trained using the EM algorithm based on a set of training label images. The average log likelihood (or bit rate) of a test set of images shows that the learned TSBN is a better model for images than models based on independent blocks of varying sizes. We also analyze the relative advantages obtained by modelling correlations at different length scales in the tree