Spatially variant mixture model for natural image segmentation

Abstract

We tackle the problem of natural image segmentation by proposing a statistical approach that is based on spatially variant finite mixture models with generalized means. The contributions can be summarized as follows: first, the proposed spatially variant mixture model exploits beta-Liouville as basic distributions for describing the underlying data structure, which demonstrated better segmentation performance than commonly used distributions, such as Gaussian; second, the mixing proportions (i.e., the probabilities of class labels) in our model are modeled via the Dirichlet compound multinomial probability density, and the spatial smoothness is imposed by adopting the function of generalized mean over the mixture model as well as mixing proportions; and finally, a variational Bayes learning approach is developed to estimate model parameters and model complexity simultaneously with closed-form solutions. The robustness, accuracy, and effectiveness of the proposed model in image segmentation are demonstrated through experiments on both natural images and synthetic images degraded by noise compared with other state-of-the-art image segmentation methods.