Abstract

A novel shift-invariant neural network (SINN) based on inter-pattern association is presented in this paper. The proposed technique is a two-step neural model, in which the binarized power spectra of a training set are calculated. Then, by using the heteroassociation, based on inter-pattern association (IPA) model, the reconnection of a shift-invariant input object can be reproduced. Since the SINN uses the IPA model, the similarities among the stored power spectra can be eliminated. Preliminary experimental demonstrations to verify the shift-invariant property of the neural network are provided.

© 1992 Optical Society of America