Noniterative manipulation of discrete energy-based models for image analysis

摘要

With emphasis on the graph structure of energy-based models devoted to image analysis, we investigate efficient procedures for sampling and inferring. We show that triangulated graphs, whom trees are simple instances of, always support causal models for which noniterative procedures can be devised to minimize the energy, to extract probabilistic descriptions, to sample from corresponding prior and posterior distributions, or to infer from local marginals. The relevance and efficiency of these procedures are illustrated for classification problems.