An Accelerated Variational Framework for Face Expression Recognition

In this paper we examine the problem of the modeling of positive vectors that are naturally generated by several signal and image processing applications. We consider a statistical framework based on a Dirichlet process mixture of inverted Dirichlet distributions like the one previously proposed in [1]. The choice of inverted Dirichlet is motivated by its flexibility to model positive features. The proposed approach can be viewed as an infinite inverted Dirichlet mixture model. As compared to [1], the infinite framework is learned via an accelerated variational approach. The accelerated variational approach offers a good compromise between frequentist techniques and purely Bayesian inference and improves the simple variational approach used in [1]. The merits of the proposed model are validated via a challenging application that involves face expression recognition.