Extracting Form Features by Bayesian Evidence Accumulation
Michael M. Marefat, Qiang Ji, Paul J. A. Lever · 1994
Abstract This paper describes an approach based on the generation and combination of geometric and topologic evidences to identify and extract semantic features from a part solid model representation [Kim 92, Finger and Safer 90]. A major difficulty faced by previously proposed methods for feature extraction has been the interaction between features. In interacting situations, the representation for various primitive features is non-unique making their recognition very difficult. The proposed method exploits Bayesian probabilistic propagation. It works by constructing and partitioning a graph which combines the original cavity graph representing the depression with a set of virtual links. However, the essence of the approach is in finding the set of correct and necessary virtual links. The evidences, which are geometric and topologic relationships at different levels of abstraction, are applied and propagated through a hierarchical singly connected hypothesis network for potential virtual links. The impact of each evidence is a measure of confidence (or probability) which causes the beliefs in all nodes of the network to be updated after arrival of that evidence. The methods for updating these beliefs are in accordance with Bayesian probability rules.