Tree-structured clustered probability models for texture

Abstract

Summary form only given, as follows. A cluster-based probability model has been found to perform extremely well at capturing the complex structures in natural textures (e.g., better than Markov random field models). Its success is mainly due to its ability to handle high dimensionality, via large conditioning neighborhoods over multiple scales, and to generalize salient characteristics from limited training data. Imposing a tree structure on this model provides not only the benefit of reducing computational complexity, but also a new benefit, the trees are mutable, allowing us to mix and match models for different sources. This flexibility is of increasing importance in emerging applications such as database retrieval for sound, image and video.