Feature extraction and shape classification of 2-D polygons using a neural network

A neural-network architecture for classification of 2-D polygonal objects is developed. The architecture is restricted to simple and viable neural mechanisms, based on those known to exist in biological neural systems. Some low-level parts of the architecture are based on the boundary contour system model of S. Grossbert (1987). The object recognition subsystem exhibits aspects of both structural/relational and decision-theoretic pattern recognition. Two key aspects of the architecture are: (1) the ability to extract and aggregate features in a hierarchical manner, such that a large number of object classes and subclasses can be recognized; and (2) the ability to transition from location-dependent feature information to location-independent feature information, such that rotational-, scale-, and translational-invariant classification is possible. Computer simulation results for one sample object are detailed.<>