A Bayesian framework for hierarchical relaxation

Our aim in this paper is to develop the formal basis for hierarchical probabilistic relaxation. The adopted approach is an evidence combining one and relies on the specification of the relaxation process in terms of Bayesian probability distributions. The approach is novel and represents a considerable advance in extending the functionality of relaxation processes. In particular, since many tasks in computer vision are formulated in terms of hierarchies of increasingly abstract image representations, the technique holds out the promise of providing a framework in which constraints from different levels can be brought to bear objectively on the interpretation task