Hierarchical MCMC Sampling
We maintain that the analysis and synthesis of random fields is much faster in a hierarchical setting. In particular, complicated long-range interactions at a fine scale become progressively more local (and therefore more efficient) at coarser levels. The key to effective coarse-scale activity is the proper model definition at those scales. This can be difficult for locally-coupled models such as Ising, but is inherent and easy for those models, commonly used in porous media, which express constraints in terms of lengths and areas.
Whereas past methods, using hierarchical random fields for image estimation and segmentation, saw only limited improvements, we find reductions in computational complexity of two or more orders of magnitude, enabling the investigation of models at much greater sizes and resolutions.
KeywordsPosterior sampling MCMC methods Hierarchical sampling Porous media Ising Random Fields Energy minimization
Unable to display preview. Download preview PDF.
- 3.Chandler, D.: Introduction to Modern Statistical Mechanics. Oxford University Press, Oxford (1987)Google Scholar
- 6.Kato, Z., Berthod, M., Zerubia, J.: A hierarchical Markov random field model. GMIP 58 (1996)Google Scholar
- 7.Puzicha, J., Buhmann, J.M.: Multiscale annealing for grouping and unsupervised texture segmentation. CVIU 76(3), 213–230 (1999)Google Scholar