A reinforced random algorithm for a partial contour perceptual similarity problem

The goal of our paper is to suggest an algorithm for the detection of contour subparts which "look similar", for example, two sufficiently large spirals twirling clockwise. We proceed in three successive stages, where in each stage we suggest some algorithm derived from the previous one and closer to the partial contour similarity (PCS) problem. We start with a simplest reinforced random algorithm, proposed for maximizing an objective function F whose arguments belong to families of G-graphs. This is achieved by a series of descents in these families, such that, as a result of a sufficiently large number of descents, attractors arise in the families. These attractors impel the following descents. The second algorithm suggested gives a rough approximation to the PCS problem. The third algorithm is obtained from the second by replacing the subsets processed by weight distributions on contours; this allows an "inexact matching" between the perceptually similar subparts sought. Experimental results are presented.