Abstract

A parametric representation of discrete image contours is presented which has the following properties: the parameters capture the perceptually relevant qualities of the contour, and the representation is compact. The descriptors are based on a segmentation of arbitrary image contours using constraints from both human visual perception and human generation of contour images. It can be proved that any such segment can be uniquely represented by a single 2-D real-valued vector. This compactness can be attributed to the use of a discrete representation and a perceptually significant segmentation. But for the parametric descriptors to be perceptually valid, a redundant representation which includes a few additional parameters is required. In addition to descriptors for individual segments, the relations between segments must be included in a complete contour description. The complete contour descriptors provide a representation which is translation, scaling, and rotation invariant for objects yet encodes relative scaling, translation, and rotation for object parts. The usefulness of the complete contour descriptors is demonstrated by their use in a neural network which learns to recognize individual handwritten digits.

© 1988 Optical Society of America