A Bayesian Information Flow Approach to Image Segmentation

Abstract

A novel Bayesian information flow approach is presented for accurate image segmentation, formulated as a maximum a posteriori (MAP) problem as per the popular Mumford-Shah (MS) model. The model is solved using an iterative Bayesian estimation approach conditioned on the flow of information within the image, where the flow is based on inter-pixel interactions and intra-region smoothness constraints. In this way, a localized and accurate Bayesian estimate of the underlying piece-wise constant regions within an image can be found, even under high noise and low contrast situations. Experimental results using 2-D images show that the proposed Bayesian information flow approach is capable of producing more accurate segmentations when compared to state-of-the-art segmentation methods, especially under scenarios with high noise levels and poor contrast.