Bayesian region merging probability for parametric image models

We describe a novel Bayesian approach to region merging, which directly uses statistical image models to determine the probability that the union of two regions is homogeneous, and does not require parameter estimation. This approach is particularly beneficial for cases in which the merging decision is most likely to be incorrect: when little information is contained in one or both of the regions and parameter estimates are unreliable. We apply the formulation to the implicit polynomial surface model for range data and texture models for intensity images.