BYY Harmony Learning of t Mixtures and Its Application to Unsupervised Image Segmentation

Bayesian Ying-Yang(BYY) harmony learning system and theory is a new kind of statistical learning approach. It has shown great advantages of parameter learning and model selection on finite mixture modeling. In this paper, we extend the BYY learning system to multivariate t mixtures and propose a gradient BYY harmony learning algorithm for t mixtures. Via optimizing the harmony function, this algorithm can determine the number of actual components during the parameter learning. Simulation experiments demonstrate that the algorithm is accurate and stable for model selection and parameter estimation. It is also robust to initializations, and suitable for model selection on high dimensional and multi-class datasets. Moreover, it is successfully applied to unsupervised image segmentation and exhibits a better segmentation performance in comparison with some typical existing algorithms.