Structural representation and BPTS learning for shape classification

In this paper, a novel shape classification technique based on a hierarchical shape representation and the back-propagation through structure (BPTS) learning algorithm is proposed. In our representation scheme, a shape is hierarchically represented with the segments composing the contour of the shape by using a scale-space filtering method. The BPTS algorithm is then applied to learn to classify shapes with such a tree-structure representation. Simulations on both artificially generated shape patterns and real world gesture patterns show that robust classification results can be achieved by using a small set of features only.