Image Segmentation Based on Finite IBL Mixture Model with a Dirichlet Compound Multinomial Prior

Abstract

In this paper, we propose a novel image segmentation approach based on finite inverted Beta-Liouville (IBL) mixture model with a Dirichlet Compound Multinomial prior. The merits of this work can be summarized as follows: 1) Our image segmentation approach is based on a finite mixture model in which each mixture component can be responsible for interpreting a particular segment within a given image; 2) We adopt IBL distribution as the basic distribution in our mixture model, which has demonstrated better modeling capabilities than Gaussian distribution for non-Gaussian data in recent research works; 3) The contextual mixing proportions (i.e., the probabilities of class labels) of our model are assumed to have a Dirichlet Compound Multinomial prior, which makes our model more robust against noise; 4) We develop a variational Bayes (VB) method that can effectively learn model parameters in closed form. The performance of the proposed image segmentation approach is compared with other related segmentation approaches to demonstrate its advantages.