Contextual decision rule for region analysis

摘要

A statistical decision rule is developed which is suitable for spatial region analysis. The rule makes use of the contextual information conveyed by triplets of objects in a neighbourhood. This rule may be used to classify objects according to a set of class labels provided that it is possible to enumerate initial noncontextual class label probabilities based on measurements for the objects, and that there exists some prior knowledge concerning the relationships between the class labels for triplets of objects. An efficient method of calculating the discriminant functions which are used to express the rule is derived. A procedure for updating the discriminant function which incorporates contextual information from outside the immediate contextual neighbourhood is described. As an example of the implementation of the contextual decision process, the problem of edge labelling is used for experimentation. The initial probabilities for a set of edge labels are obtained using two 3×3 edge masks. The a priori probability of label triplets is inferred from a set of syntactic rules for edge continuity which embody local space symmetries.