A Hierarchical Framework for Shape Recognition Using Articulated Shape Mixtures

Abstract

This paper describes a statistical framework for recognising 2D shapes with articulated components. The shapes are represented using both geometrical and a symbolic primitives, that are encapsulated in a three layer hierarchical architecture. Each primitive is modelled so as to allow a degree of articulated freedom using a polar point distribution model that captures how the primitive movement varies over a training set. Each segment is assigned a symbolic label to distinguish its identity, and the overall shape is represented by a configuration of labels. We demonstrate how both the point-distribution model and the symbolic labels can be combined to perform recognition using the hierarchical mixture of experts algorithm. This involves recovering the parameters of the point distribution model that minimise an alignment error, and recovering symbol configurations that minimise a structural error. We apply the recognition strategy on sets of Arabic characters.