Identification of a generic MGRF model involves estimating both the characteristic neighbourhood and the corresponding potentials. Conventionally, the identification involves computing the analytical first approximation of potentials, deriving characteristic stature, and refining potentials via stochastic approximation [38].
In this thesis, a structural analysis of stochastic and regular (nearly periodic) image textures is considered, which focuses on estimating arbitrary shaped texture elements (textons [54] or texels [43]) and rules of their spatial placement for texture description. The proposed method derives the geometric structure and the placement rule for texels, from an analytically identified MGRF model, via spatial analysis of the patterns formed by characteristic clique families in a model-based interaction map. The structural identification is computationally much more efficient compared to the identification via MCMC algorithms. In addition, the resulting texel-based texture description leads to a fast technique for realistic texture synthesis-via-analysis, namely, bunch sampling.
As a whole, the main contribution of this thesis is in demonstrating a hybrid method that enables fast texture analysis and synthesis by combining the complementary strength of both probability model-based and structural approaches.