Unsupervised color images segmentation using spatial hidden MRF GDPM model

Abstract

Model based methods of segmentation have largely provided a convenient format framework for segmenting observed color images. But most of these models assume that the number of components in mixture model is known in advance. To resolve this issue, we proposed a novel variational Bayesian segmentation method using a Spatial Hidden Markov Random Field Gaussian Dirichlet Process Mixture (Spatial-HMRF-GDPM) model to segment automatically given color images. In particular, our algorithm has used the concept of Spatial Hidden Markov Random Field to model the spatial correlation structure between the latent labeling variables of observed data points in color space, and has used the Gaussian Dirichlet Process Mixture model to represent a likelihood function with infinite number of class memberships for each observation. We have also described how to use an efficient variational Bayesian algorithm to inference the proposed model. Finally, we apply this model to segment a series of various color images, demonstrating its advantages over existing methodologies. Experimental results show that our algorithm manage to discriminate exactly each objects in color images, and also the noise in the segmented image is significantly reduced. Moreover, the convergence speed of our method is very faster than other methods.