A systematic method for exploring contour segment descriptions

Abstract

A new methodology has been developed for the comparative analysis of different contour characterization methods in the context of a structural approach for an artificial vision system. This methodology has been designed for testing the description of contour segments using characterization methods that have been used in applications dealing with isolated forms. The proposed schema is modular and comprises three major steps: a segmentation stage, which is based on a rule-based segmentation machine; a labeling stage, which is dependent on the characterization method used to describe the contour segments and in which the problem of defining classes or typologies of segments is also considered; and a learning and recognition stage, which uses fast tree classifiers and combines their evidence by means of the Dempster-Shafer combination rule. Experimental results are presented using the Fourier-Bessel transform as the contour characterization method.