Image analysis that produces an image-like array of symbolic or numerical elements (such as edge finding or depth map reconstruction) can be formulated as a labeling problem in which each element is to be assigned a label from a discrete or continuous label set. This formulation lends itself to algorithms, based on Bayesian probability theory, that support the combination of disparate sources of information, including prior knowledge. In the approach described here, local visual observations for each entity to be labeled (e.g., edge sites, pixels, elements in a depth array) yield label likelihoods. Likelihoods from several sources are combined consistently in abstraction-hierarchical label structures using a new, computationally simple procedure. The resulting label likelihoods are combined with a priori spatial knowledge encoded in a Markov random field (MRF). From the prior probabilities and the evidence-dependent combined likelihoods, the a posteriori distribution of the labelings is derived using Bayes' theorem. A new inference method, Highest Confidence First (HCF) estimation, is used to infer a unique labeling from the a posteriori distribution that is consistent with both prior knowledge and evidence. HCF compares favorably to previous techniques, all equivalent to some form of energy minimization or optimization, in finding a good MRF labeling. HCF is computationally efficient and predictable and produces better answers (lower energies) while exhibiting graceful degradation under noise and least commitment under inaccurate models. The technique generalizes to higher-level vision problems and other domains, and is demonstrated on problems of intensity edge detection and surface depth reconstruction. © 1990 Kluwer Academic Publishers.