This PDF file contains the front matter associated with SPIE Proceedings Volume 10262, including the Title Page, Copyright information, Table of Contents, and Conference Committee listing.

Efficient signal processing is made possible by using hybrid electro-optical systems which can be trained to perform a specific task. Various architectures and procedures are implemented for optimization of pattern recognition and other signal processing systems.

This paper reviews work on binary phase-only (BPOF) and ternary phase-amplitude (TPAF) correlation and highlights recent investigations of neural network approaches for augmenting correlation-based hybrid (optical/electronic) automatic target recognition systems. The theory and implementation of BPOF and TPAF correlation using available spatial light modulators is reviewed, including recent advances in smart TPAF formulations. Results showing the promise of neural networks for enhancing correlation system operation in the areas of estimating distortion parameters, adapting filters, and improving discrimination are presented and discussed.

Optical correlation schemes based on a matched filter containing a single Mellin radial or logarithmic harmonic of an object can be used to provide 2-D or 1-D scale invariant pattern recognition. The performance of these harmonic filters, as well as the performance of the conventional matched filter, was investigated. The response of the filters to various distortion parameters was studied along with a quantitative comparison between the various filters, using several performance criteria. In addition, the noise response of the above filters is analyzed. The noise considered was an additive white noise to the input plane, characterized by a single parameter σ, varied between 0 to 3 times the maximum input intensity. The analysis and results apply also to the joint transform correlator configuration.

This paper reviews recent progress in rotation invariant pattern recognition; the emphasis is on the work done in our own laboratories, since much of the significant work done elsewhere is described in other papers presented at this conference.

Various optical feature extraction and spatial filters for scale, rotation and shift invariant pattern recognition are reviewed. Many methods are based on the circular Fourier and radial Mellin transform (FMT). With the pure imaginary Mellin transform order s=-jω, the FMT may be implemented using a polar-log coordinate transform followed by 2D Fourier transform. With positive integer orders s, the FMT’s yield the moment invariants. Invariant pattern recognition is made using the regular moments for image normalization, the Hu’s moment invariants, the Zemike moment invariants, the complex moments, the Fourier-Mellin descriptors, and the orthogonal Fourier-Mellin moments. The polar coordinate Fourier-Mellin moments use more low order moments and do not suffer from information suppression, information redundancy and are more robust against noise. Optically generated image moments may be combined to the moment invariants. The FMT may also be directly generated in an optical correlator using the Fourier-Mellin spatial filters (FMF’s), that allow an additional shift invariance. An optoneural system using the FMF and a neural network is presented.

Optical processing systems have been in existence for over 20 years but few have made the transition from the laboratory environment to the commercial environment. Pattern recognition done with optical correlators is one area which is now showing promise for applications outside the laboratory. Many papers have been written in years past on the need for better and faster spatial light modulators and more efficient filtering schemes prior to commercializing an optical correlator. The current state-of-the-art is now approaching this point. Two companies are even marketing Liquid Crystal Television based correlators. The purpose of this review will be to examine the current capabilities of optical correlators and to match those capabilities to existing requirements.

Various nonlinear mapping techniques have been investigated for exploiting the restrictive characteristics of available spatial light modulators (e.g., phase-only or binary phase-only) and/or to modify the performance of coherent optical correlators. The phasewith- constrained-magnitude filter (PCMF) is a practical complex spatial filter realizable with today’s technology. It relies on the coupling between magnitude and phase transmittance exhibited by certain types of spatial light modulators. It requires relatively little computational effort to implement, and its performance has been shown to compare quite favorably with that of phase-only filters. In this paper we review the basic PCMF concept and show how performance of the PCMF can be improved by modifying the basic PCMF algorithm in two ways: (1) by setting the magnitude transmittance of the spatial light modulator to its lowest achievable value for a chosen set of low spatial frequency components and (2) by adjusting the values of the desired spatial frequency distribution relative to the available dynamic range of the spatial light modulator. The second method, in particular, can be used to make the PCMF perform more like an inverse filter or more like a matched filter, depending on the desired results.

We review the nonlinear joint transform correlator (JTC) including the kth law nonlinear JTC, and the binary nonlinear JTC. Multiple objects detection using binary joint transform correlation are presented. An optimum threshold function is computed to maximize the light intensity of the correlation peak and to eliminate the even order harmonic terms. The correlation performance of the binary JTC is determined for different thresholding methods. The binary JTC produces large peak to sidelobe ratio and narrow peak for the multiple targets. The optimum threshold function produces better correlation performance compared with the median thresholding.

An overview of the field of optical pattern recognition using photorefractive media is presented. The topics include spatial light modulation, parallel image subtraction, parallel spectrum subtraction, and real time optical correlation.

Photorefraction is a nonlinear optical effect where large nonlinearities can be observed at relatively low optical power levels. This phenomenon has evolved from being just a laboratory curiosity to being a useful device technology for optical information processing applications. In particular, the abilities to form holograms in real time and also store them for extended periods of time have aided the implementation of several pattern recognition machines which are adaptive and possess a high degree of parallelism. This paper reviews the photorefractive effect as a device technology for the problem of pattern recognition and describes several photorefractive realizations of pattern classifiers.

Designing filters for use with optical correlators is really an exercise in trading one performance measure against another. In this critical review, we present several different situations where such a tradeoff is carried out. An informed understanding of this law of nature is important in making sure that our goals in optical pattern recognition are realistic.

Until recently, optimal filter theory has not explicitly included several limitations imposed by the physical nature of spatial light modulators and detectors. We review recent theory that includes colored input noise; finite magnitude and finite contrast ratio of a spatial light modulator; single-drive spatial modulators that couple their effect between amplitude and phase, or are limited to real values; and electronic noise in correlation detection. We set up a signal to noise metric to optimize, and we use several approaches to its optimization. One is the Cauchy-Schwartz inequality, another is differential calculus (variational calculus and partial differential equations), and the third is a relaxation similar to annealing.

In this paper, we review correlation filters as an approach to pattern recognition with a special emphasis on the consequences of normalizing the correlation to achieve intensity invariance. Intensity invariance is effected using the Cauchy-Schwarz inequality to normalize the correlation integral. We discuss the implications of this criterion for the application of correlation filters to the pattern recognition problem. It is shown that normalized phase-only and synthetic discriminate functions do not provide the recognition/discrimination obtained with the classical matched filter.

Nonlinearly transformed matched filters including the k^th law nonlinear filters for optical correlation are reviewed. The interesting property of the k^th law nonlinear filters is that the nonlinear transformation modifies the amplitude of the conventional matched filter by a power of k. Various types of filters may be produced by varying the severity of the nonlinearity. Correlation results for the images tested indicate that nonlinear matched filters produce good correlation performance in the terms of correlation peak intensity, signal-to-noise ratio, and peak-to-sidelobe ratio. The sensitivity of the k^th law nonlinearly transformed filter to rotational changes of the input signal is investigated. The rotational sensitivity of the k^th law nonlinearly transformed filter increases for higher degrees of nonlinearity. However, for a given input signal rotation, a highly k^th law nonlinearly transformed filter could produce a better peak-to-sidelobe ratio and better SNR compared with a moderately k^th law nonlinearly transformed filter and/or a conventional matched filter.