Abstract

A complete learning, and pattern-recognition system based on the joint-transform correlator is presented. By employing an iterative process, a complex synthetic-discriminant function (SDF) for pattern classification is generated, which takes into account all the constraints and distortions of the actual system. To effectively implement a complex reference function with a binary mask, the following steps are taken. The input pattern, multiplied by a grating, is introduced into the correlator at some distance from a binary reference function. After an optical Fourier-transform is performed on the input plane, a region around the first diffraction order is recorded by a television camera and is displayed on a spatial light modulator (SLM). The correlation is achieved by a second optical Fourier-transformation. The classification criterion is a high peak for one class of a training set and a uniform distribution for a second class. The elements of the SDF are the variables of a cost function, which is minimized by an iterative method to achieve this recognition criterion. All Fourier transforms are performed optically, but the cost function is calculated on a digital computer. Experimental results demonstrate high quality performance even with a poor SLM.

© 1990 Optical Society of America