Nonparametric hierarchical mixture models based on asymmetric Gaussian distribution

Abstract

Mixture models are broadly applied in image processing domains with hierarchical Bayesian approaches popularly considered for grouped data. The related inference process takes into consideration both the approximation of exact data shapes and estimation of adequate component numbers. In our work, we develop nonparametric hierarchical Bayesian models using the Dirichlet and Pitman-Yor processes with asymmetric Gaussian distribution. The parameters of these models are learned using variational inference methods. The effectiveness and merits of the proposed approaches are validated using the challenging real-life application of dynamic texture clustering.