Abstract
In this paper, we focus on constructing new flexible and powerful parametric framework for visual data modeling and reconstruction. In particular, we propose a Bayesian density estimation method based upon mixtures of scaled Dirichlet distributions. The consideration of Bayesian learning is interesting in several respects. It allows simultaneous parameters estimation and model selection, it permits also taking uncertainty into account by introducing prior information about the parameters and it allows overcoming learning problems related to overor under-fitting. In this work, three key issues related to the Bayesian mixture learning are addressed which are the choice of prior distributions, the estimation of the parameters, and the selection of the number of components. Moreover, a principled Metropolis-within-Gibbs sampler algorithm for scaled Dirichlet mixtures is developed. Finally, the proposed Bayesian framework is tested on a challenging real-life application namely visual scene reconstruction.