Image Segmentation Using Markov Random Field Combined with Hierarchical Prior Models

Abstract

Image segmentation can be viewed as a pixel classification problem, where one would assign a label to each pixel within an image in order to specify the class to which the pixel should belong. As with the estimation of labels, it is not easy to determine optimal values for model parameters. Our model proposed in this work could be effective in cases of the lack of sufficient information by introducing to the priors of parameters a hierarchical structure that would make minimal assumptions on the data. Markov chains in Gibbs sampling would efficiently explore the joint posterior of segmentation labels and model parameters. We can use samples obtained by simulating this chain to perform Bayesian inference of labels. We attempt a quantitative evaluation of image segmentation algorithms by comparing their results to those given by humans. © 2006, The Institute of Image Electronics Engineers of Japan. All rights reserved.