Bayesian Colour Image Segmentation using Pixon and Adaptive Spatial Finite Mixture Model

A colour image segmentation using pixon-representation and a modified finite mixture model called Adaptive Spatial Finite Mixture Model is proposed in this paper. First the colour image is described using fewer degrees of freedoms where the characteristics of colour image are homogeneous and using more degrees of freedom where there is heterogeneity, which is called pixonrepresentation. An adaptive thresholding technique which uses quaternion moments is used to capture the regional homogeneity. Then, these image pixons are replaced by their corresponding features and adjacencies, thus forming a pixon-map of the observed image. The key idea to this approach is that this pixon map is embedded into an adaptive spatial finite mixture model under a Bayesian framework. Experimental results with Berkeley segmentation dataset, Corel database images and some natural images illustrate that the proposed method is much more effective and powerful in colour image segmentation than the other pixon-based approaches.