A split and merge EM algorithm for color image segmentation

As an extremely powerful probability model, Gaussian mixture model (GMM) has been widely used in the fields of pattern recognition, information processing and data mining. However, in many practical applications, the number of the components is unknown. In the case, model selection of GMM, i.e., the selection of the number of the components in the mixture, has been a rather difficult problem. Recently, the minimum message length (MML) criterion has been proposed and used to make model selection. In this paper, we propose a split and merge algorithm to decide the number of the components, which is applied to the color image segmentation. Based on MML criterion, the proposed algorithm can determine the number of components in the Gaussian mixture model automatically during the parameter learning. By splitting and merging the incorrect components, the algorithm can converge to the maximization of the MML criterion function and get a better parameter estimation of the Gaussian mixture. It has been demonstrated well by the experiments that the proposed split and merge algorithm can make both parameter learning and model selection efficiently for color image segmentation.