Multi-resolution shape description and recognition from 2-D silhouette based on tree representations of skeletons

Abstract

A shape description and recognition algorithm is proposed. The description algorithm generates a hierarchical structural description of shape automatically from a two-dimensional silhouette. The description is based on skeletons, which are very useful for describing nonrigid bodies such as animals. Moreover, this hierarchical description includes not only quantitative features at each resolution level, but also the relationship between hierarchical levels by describing a skeleton as a tree representation. An effective matching method is also proposed which can discriminate shapes hierarchically using a three-layered neural network learned by an error back-propagation rule.<>