Neural network architecture for circular features extraction in binary patterns

A new neural network architecture developed for circular features recognition in binary images is introduced. The methodology involves a so called ´growing the field of vision´ technique and employs a new transfer function based on the classical sigmoid function. At any instant the method processes only cells within a circular cluster and as time progresses the network searches for circular features of greater radii. The method is immune to translation, scaling, rotation, and, depending on the training schedule, distortion of patterns. Training procedures are presented together with the results of the recognition of some example patterns.