Learning the Structure of Textures

A framework for learning the structure of textures is introduced in this work. A Bayesian network, whose structure and parameters are learned from a small texture sample, is used to characterize the high-dimensional probability density function (PDF) that models the given texture. The underlying network learning framework is based on the combination of tools such as kernel density estimation and quasi-greedy tree search, together with optimality criteria such as maximum likelihood and crossvalidation. While learning the Bayesian network, in addition to the PDF, we learn critical structural relationships among the texture pixels. As examples of the proposed framework, we present two basic applications: First, the learned PDFs are used for texture classification, through a Bayes classifier; and second, the structural knowledge is used to improve state-of-the-art texture synthesis algorithms by customizing the context to each texture.