Abstract

The associative content-addressable memory proposed by Hopfield has a good tolerance for errors and is capable of solving image recognition problems. The error correction capability for such a model of size N is limited to (N − 1)/2 bits, which is just under 50%. Our proposed M-ary network can achieve an error correction capability of (N − 1)(M − 1)/M digits, with a theoretical upper limit of (N − 1) digits, one digit short of 100%. This Substantial improvement comes at the cost of increasing the complexity of the feedback thresholding operation. An electroptical implementation of the model is described. The first component, an optical preprocessor, performs a Fourier transformation of the optical signal. This increases the tolerance for size and orientation discrepancies of the object. The second component is a CCD camera which transforms the 2-D optical signal into an analog electrical signal. The third component, a frame grabber, digitizes and stores the analog signal. The fourth component is a 16-bit Zoran digital signal processor (DSP) which performs the necessary calculations and the feedback thresholding operations required. The DSP under consideration would process forty images. Each image consists of 1024 M-ary digits (32 × 32 pixels). The processed signal is then projected onto a video monitor.

© 1988 Optical Society of America