Texture analysis and synthesis have been two fundamental research subjects in computer vision because of an increasing variety of potential applications being created. The primary focus of this thesis is on developing a structural identification of a generic Markov-Gibbs random field model of textures which results in a new approach for fast texture analysis and synthesis.
Probability models, in particular Markov-Gibbs random fields (MGRF) have gained wide acceptance for solving applied image recognition, analysis, and synthesis problems. An MGRF model of textures is usually specified by a Gibbs probability distribution on selected image features such as image signal statistics. Despite of modelling power and expressiveness of an MGRF model, identification of the model, i.e. estimating unknown parameters to learn a probability model for an image, is computationally complex because the process usually involves Markov Chain Monte Carlo (MCMC) algorithms with exponential time complexity.
This thesis considers a structural analysis of spatially homogeneous stochastic and regular textures. The proposed method describes a texture by identifying its characteristic texels (texture elements) and deriving placement rules of their spatial organisation. The identification conducts a structural analysis of a model-based interaction map obtained from an analytically identified generic MGRF model. As a result, the method provides fast model identification, and meanwhile, leads to a novel texel-based texture description and a fast texture synthesis algorithm.
Results of this work have been reported in [40,108,109,110,111].