Multiresolution Approach for Color Image Segmentation using MRF Model

Abstract

In this paper, the color image segmentation problem is addressed in supervised framework. In the supervised framework, we assume to have one original image from the class of images from which the given image is derived. In this framework, We have used Markov Random Field(MRF) to model the image label process and the MRF model parameters are estimated using the conditional pseudolikelihood criterion. Ohta( I1, I2, I3 )model is used as the color model. The segmentation problem is formulated as the pixel labeling problem. The image model parameter estimation problem is formulated using pseudo-likelihood criterion. The image label estimation problem is cast in Maximum {a Posteriori} (MAP) framework. These MAP estimates are obtained using the proposed new hybrid algorithm and compared with Simulated Annealing (SA) algorithm. It is observed that the proposed hybrid algorithm converge much faster than that of SA. The segmentation scheme is further improved by adhering to multiresolution approach where the segmentation is carried at a coarse level and reconstructed to a finer level. The proposed scheme appears to be more viable from a practical standpoint.