Invariant object recognition using a neural template classifier

This paper describes an efficient two-stage neural network for invariant object recognition. It consists of a feature extractor trained by an ART-like fast saturation learning scheme and a delta-rule trained classifier. Objects, represented as edge strength maps derived from raw input images, are scaled to a normalized size and rotated in discrete steps to generate a sequence of localized input feature vectors. The network outputs identify the object and permit the calculation of a confidence level. Experiments show that the system works well even when there is noise and occlusion.