From discs to parts of visual form

摘要

Object recognition can be favoured by the development of computational models leading to reliable descriptions of the shape of bidimensional patterns. We refer to the approach regarding a pattern as made up by a number of parts, and present a partition procedure where the parts are obtained by merging elementary regions grown from feature sets found on the Distance Transform. The feature sets include most of the centres of maximal discs, and are placed in correspondence with pattern subsets that can be qualitatively described as near-convex regions, protrusions and necks. The pattern partition found appears adequate to describe the pictorial data in a natural manner that does not lack psychological reality.