Color Image Segmentation Using Generalized Inverted Dirichlet Finite Mixture Models By Integrating Spatial Information

Abstract

Mixture models are popular statistical approaches for image segmentation. However, mixture models based segmentation faces some difficulties. The first problem is the estimation of the number of clusters (M). Secondly, the spatial information is generally not considered. In this paper, we have used two methods to counter these issues. The first method uses spatial information as the prior knowledge of M. This prior knowledge does not give the direct value of M instead it provides some indirect information which can be used to estimate the optimal value of M. The second one uses Markov Random Field (MRF) to integrate spatial information. MRF based models need high computational power due to their complexity. They cannot be used directly in the Maximization step (M-Step) of Expectation-Maximization (EM) algorithm. The MRF model used in this paper does not require high computational power and can be easily integrated with the M-Step. We have implemented Inverted Dirichlet (ID) and Generalized Inverted Dirichlet (GID) mixture models using these two methods. For experiments, we have used 500 Berkeley dataset (BSD500). In order to compare the image segmentation results, the outputs of ID mixture model (IDMM) and GID mixture model (GIDMM) are compared with the Gaussian mixture model (GMM), using segmentation performance evaluation metrics. The results obtained from GIDMM and IDMM are more promising than those obtained with GMM.