Encoding contour shape by curvature extrema

Abstract

Curvature extrema provide significant information about the shape of an image contour, such as a silhouette, and are the basis for the Hoffman-Richards codon representation for shape. This representation based on curvature easily translates into a binary string that will describe the abstract shape of any smooth image curve. The computation of the basic shape primitives requires dealing with two ever-pervasive problems: contour noise and scale. We show how contour noise can be estimated given knowledge of the shape of the filter used to compute curvature from the edge list of the contour. To handle the scale problem, we use an adaptation of Witkin's scale space. Our algorithm differs from Witkin's by using a notion of parts to set criteria for significant structures.