Unsupervised learning of a finite discrete mixture: Applications to texture modeling and image databases summarization

This paper presents an unsupervised learning algorithm for fitting a finite mixture model based on the Multinomial Dirichlet distribution (MDD). This mixture is particularly useful for modeling discrete data (vectors of counts). The algorithm proposed is based on the expectation maximization (EM) approach. This mixture is used to improve image databases categorization by integrating semantic features and to produce a new texture model. For the texture modeling problem, the results are reported on the Vistex texture image database from the MIT Media Lab.