This article addresses the development of and recent advances in the rapidly growing field of optical pattern recognition. In optical pattern recognition there are two basic approaches; namely, matched filtering and associative memories. The first employs optical correlator architectures and the latter uses optical neural networks (NNs). This paper reviews various types of optical correlators and NNs applied to real-time pattern recognition and autonomous tracking. Techniques of scale and rotational invariant filtering are also given. Recent approaches using wavelet transform filtering, phase only filtering, high capacity composite filters and phase representation for improvement in pattern discrimination are also provided.

An inter-pattern association (IPA) neural network model is presented in which basic logical operations are used to determine the association among common and special features of reference patterns. Hetero- and auto-associative memory are synthesized by applying a generalized logical rule. Computer simulations for pattern recognition by using the IPA model have shown a better performance and a higher storage capacity than the Hopfield model. A 2-D adaptive optical neural network is used to perform parallel neurocomputations. Since the interconnection weight matrix for the IPA model has a tristate structure, the dynamic range imposed on a spatial light modulator is rather relaxed, and the interconnections are much simpler than for the Hopfield model.>

Generally, the synthesis of coherent spatial filters is restricted to the linear region of the transfer characteristic of a photographic film. However, a technique of synthesizing a nonlinear spatial filter such that the signal detection may be optimum will be described. In this paper, a generalized linear optimization technique is formulated. The application of this optimization technique toward a simple nonlinear spatial filter is demonstrated, and the extension of this optimization technique for a more complicated nonlinear spatial filter is also given. The signal detection by nonlinear optimum spatial filtering is analyzed. Finally, it is concluded that, instead of restricting the spatial-filter recording to the linear region of the transfer characteristic of the photographic film, an optimum nonlinear spatial filter may be achieved.

A different approach toward the information channel capacity of a photographic film is presented. It is based on the total number of possible combinations of the smallest distinguishable regions in a piece of photographic film. The information channel capacity of a hologram is also discussed. It is found that the information capacity of a hologram is somewhat lower than that of a photographic film. However, the information content of a hologram is, in general, greater than that of a photograph. It is the recording process used in making a hologram that provides the advantage over an ordinary photograph. Nevertheless, if the object to be recorded were properly encoded, then the information content of a hologram could never reach that of a photographic film.