A generic MGRF model involves an arbitrary structure of multiple pairwise pixel interactions, which is specified by a Gibbs distribution with the Gibbs energy being a linear combination of clique potentials as model parameters and GLCH statistics as sufficient statistics. The model distribution belongs to exponential families and is also an MaxEnt distribution.
Given a training image, the traditional model identification of a generic MGRF involves deriving the first approximation of potentials, identifying the characteristic neighbourhood and refining clique potentials via stochastic approximation. The identification is also considered as a process of texture synthesis.
The main motivation of the proposed structural identification is to simplify the traditional computation-intensive model identification via stochastic approximation. Structural identification involves estimating the geometric structure and placement rules of texels, from an analytically identified MGRF model, in particular by a structural analysis of spatial patterns formed by clique families in an MBIM. The identification results in a texel-based texture description and a fast texture synthesis algorithm.