Complex associative memory neural network model for invariant pattern recognition

Abstract

A complex associative memory neural network (CAMN/sup 2/) model is proposed for the recognition of handwritten characters. The input and the stored patterns are derived from the complex valued representation of the boundary of the characters. The stored vector representation is formulated based on 1-D representation of an optical pattern recognition filter. Retrieval of stored patterns from a noisy and shifted input is accomplished by using the correlation in the inverse Fourier domain. An adaptive thresholding scheme is then applied to obtain a 1-step convergence. The number of convergence of patterns, usually measured as the storage capacity of the associative memory is found to increase significantly. But the major advantage obtained from the complex representation is that the recognition of patterns is invariant to translation, rotation and scaling of the input patterns.<>