For simplicity, the structural identification assumes a uniform geometric structure for all texels in a texture, i.e. the same pixel interactions are applicable at every location in the image. The geometric structure is statistically the most probable one, i.e. the central tendency if consider a distribution of all possible texel's structures. This simplification is implied by the assumption of translational-invariant pixel interaction in a generic MGRF model.
The geometric structure of texels in a texture is derived from the characteristic neighbourhood consisting of local maxima of clusters. Structures for regular and stochastic textures are specified in separate ways, due to the different nature of two texture classes.