A method for the analysis of ambiguous segmentations of images

In this correspondence we are interested in how interpretation and context restrictions can guide the analysis of ambiguous segmentations of images in computer vision systems. The final objective is to find image segments that can be interpreted (classified) such that their interpretations do not conflict with interpretations given to related segments. In the case that we have several possible labels for each segment, some of this ambiguity can be reduced by means of a relaxation process. In its discrete formulation, the relaxation operator examines pairs of related segments to see if they have incompatible labels, which are then discarded. This process is iterated until only compatible labels are left. In this work a new approach is proposed that considers all possible segmentations resulting from an ambiguous segmentation simultaneously in only one relaxation process. A new relaxation operator is defined that can be applied to ambiguous segmentations. In this way no backtracking is performed, ambiguity is reduced, and the best solution is still retained. The output of the process is a collection of segmentations and interpretations that is hopefully small enough so that each case can be considered separately.