Abstract

An optical neural network that is based on the neocognitron paradigm proposed by K. Fukushima is implemented and experimentally demonstrated for automatic target recognition. A novel aspect of the optical architectural design is the use of a shift-invariant multichannel optical correlator within each processing layer. Multilayer processing is achieved by iteratively feeding back the output of the feature correlator to the input spatial light modulator and updating the Fourier filters. Characteristic features are first extracted from the object images as the training set. Neural synaptic weights are stored in Fourier hologram array. Distortion invariance is greatly improved with the use of these training features. As an example of experimental demonstration, simulated reentry vehicle and decoys image are used as the test objects. Successful recognition and classification of these objects with intraclass fault tolerance and interclass discrimination is achieved. A detailed system description and experimental demonstration are presented.

© 1992 Optical Society of America