Motivation

Texture analysis is an important research area in computer vision. Enormous efforts have been made in search of an efficient texture description. However, texture analysis is generally a difficult problem due to diversity and complexity of natural textures.

Probability models have gained wide acceptance, because of their modelling power and expressiveness. These models pose the problem of texture analysis into a statistical setting, which allows a wide range of well-established theories and methodologies in mathematical statistics to be introduced into texture modelling. In particular, Markov-Gibbs random fields (MGRF), which describe a texture in terms of spatial geometry and quantitative strengths of inter-pixel statistical dependency, are among the most successful probability models.

A probability model is usually specified by a parametric probability distribution. The model is to be `identified', in order to find best values for unknown parameters of the model for a given training texture. Due to usually complex mathematical form of the distribution, model identification is not trivial. For instance, identification of a generic MGRF model involves a process of stochastic approximation having exponential time complexity [37].

This thesis considers a more efficient identification of a generic MGRF model that characterises a texture by a group of texels and placement rules of their spatial arrangement. Combining the strength of model-based analysis and structural approach, the proposed method results in a novel texel-based texture description and leads to a technique for fast texture synthesis.