Summary

The bunch sampling is a texture synthesis technique derived from a structural identification of a generic MGRF model. In the synthesis, the geometric structure and the placement rules for texels, both resulted from the identification, are used for sampling texels from a training image and guiding the placement of obtained texels onto a synthetic texture, respectively. Experimental results show that the bunch sampling captures and replicates visual characteristics in a broad range of homogeneous textures, both regular and stochastic.

In the bunch sampling, the time complexity of the analysis and the synthesis stages are $O({\vert{g}^{\circ}\vert}^2)$ and $O({\vert{g}^{syn}\vert})$ respectively. So, the proposed method is very fast compared to the model-based synthesis with exponential time complexity and the synthesis based on non-parametric sampling involving exhaustive search for neighbourhood matching.

However, the bunch sampling has a major limitation. Based on global signal statistics for feature description, the bunch sampling is not effective in capturing local deformation exhibited in many weakly-homogeneous textures. Although local deformation is revealed by the shape and the size of clusters in an MBIM, the method assumes that a same structure of texel is applicable at every pixel in an image and derives a structure based on an approximate Bayesian estimate of texels. The oversimplified approach neglects all information about local deformation, and, as the experimental results show, the synthesis only produces idealised textures by using a single, the most probably texel structure. It remains still an open problem to cope with dynamic geometric structures accounting for variability of texels at different image locations.