Abstract

Character recognition by a trinary associative memory (TAM) neural network model is proposed. All the twenty-six letters of the English alphabet are stored in the trinary memory. The proposed scheme will then be able to recognize a character from its partial input. The dot product of the partial input with all the stored patterns is calculated as a measure of discrepancy from the desired pattern. Zero thresholding and arithmetic mean thresholding and some other statistical thresholding methods are then applied to select the desired output. The convergence of the recall procedure of the TAM network depends upon the storage representation and thresholding mode. So from the simulation run, an optimum threshold measure is found that establishes efficient character recognition. Shift invariance is also examined by simulation.

© 1991 Optical Society of America