Neural networks employing pulsating biology-oriented integrate-and-fire (IF) model neurons, that can exhibit synchronicity (phase-locking), bifurcation, and chaos, have features that make them potentially useful for learning and recognition of spatio-temporal patterns, generation of complex motor control, emulating higher-level cortical functions like feature binding, separation of object from background, cognition and other higher-level functions; all of which are beyond the ready reach of nonpulsating sigmoidal neuron networks. The spiking nature of biology-oriented neural networks makes their study in digital hardware impractical. Prange and Klar convincingly argued that the best way of realizing such networks is through analog CMOS technology rather than digital hardware. They showed, however, that the number of neurons one can accommodate on a VLSI chip limited to a hundred or so, even when submicron CMOS technology is used, because of the relatively large size of the neuron/dendrite cell. One way of reducing the size of neuron/dendrite cell is to reduce the structural complexity of the cell by realizing some of the processes needed in the cell's operation externally to the chip and by coupling these processes to the cell optically. Two such processes are the relaxation mechanism of the IF neuron and dendritic-tree processing. We have shown, by examining the blue light impulse response of electron trapping materials (ETMs) used under simultaneous infrared and blue light bias, that these materials offer features that can be used in realizing both the optical relaxation and synapto-dendritic response mechanisms. Experimental results demonstrating the potential of this approach in realizing dense arrays of biology-oriented neuron/dendrite cells will be presented, focusing on the concept and design of ETM-based image intensifier as new enabling technology.

We describe a laser diode-based optoelectronic implementation of artificial neural networks which utilizes real-time holography in photorefractive crystals. The use of a laser diode light source reduces the system size and power requirements. The holographic material is rhodium- doped BaTiO₃ which has enhanced sensitivity at the laser-diode wavelength of 830 nm. A balanced coherent-detection method is used to represent bipolar optical neurons and weights. In addition, by distributing each neuron weight among a set of spatially and angularly distributed gratings using beam fanning, Bragg degeneracy and its associated inter-neuron optical crosstalk is virtually eliminated. The structure of the neural network is programmable and we have implemented a variety of neural networks including backpropagation and Kohonen-style self-organizing maps with up to 10,000 neurons and performance of up to 10⁸ weights processed per second during learning and readout. We also discuss weight decay in photorefractive materials, specifically its relative effects in the neural network and data storage domains.

We introduce a new concept called a pulse-coupled neural network (PCNN). This system works by inputting a two dimensional image into the PCNN. The PCNN iterates over a set of five equations generating an output of summed pulses versus time. The output is solely dependent on the shape and the intensity of the image input into the network.

We describe a nonlinear joint transform correlator-based two-layer network that uses a supervised learning algorithm for real-time face recognition. The system is trained with a sequence of facial images and is able to classify an input face image in real-time. Computer simulations and optical experimental results are presented. The processor can be manufactured into a compact low-cost optoelectronic system. The use of the nonlinear joint transform correlator provides good noise robustness and good image discrimination.

A programmable multilayer neural network is under construction to illustrate the advantages of optical interconnects (scaleability, bandwidth) and to overcome the disadvantages of optical devices (limited precision, nonideal transfer characteristics, lack of subtraction). The disadvantages will be surmounted with the help of dedicated software engineering and additional optical hardware. The complete system design is summarized, and the construction of the first matrix vector multiplier subunit is detailed together with first test results. Finally, the progress made on the remaining subunits is reported with a perspective to future work.

The use of multiple eigenvectors is proposed for improving the performance of multiclass distance classifier correlation filters (DCCFs). The results of tests to evaluate the performance of an automatic target recognition (ATR) system employing DCCFs as a function of range and resolution limits are presented. The results obtained demonstrate the ability of the proposed ATR to perform discrimination between similar targets at long ranges, and to successfully locate and identify targets imbedded in realistic background and clutter.

We analyze the statistical properties of a new optical encoding method of images for security applications. The encoded image is obtained by random phase encoding in both input and Fourier planes. We show that this encoding method converts the input signal to a stationary white noise and that the reconstruction method is very robust.

A method is investigated for fast online inspection of pins on very large scale integrated (VLSI) circuit chips prior to mounting on circuit boards. We image the edge of the VLSI chip containing the pins to be inspected onto electron trapping material (ETM) using Argon light. This writes an image of the pins to be inspected onto the material. Then we image an identical chip type with perfect pins onto the electron trapping material using IR light. This performs a subtraction between the two images. The output is read out with another IR beam and summed onto a detector. If the output exceeds an electronic threshold, pins are considered damaged. We show experimental results with pins separated by 0.2 mm to show practicality and speed of the approach. A second approach is explored in which the grating effect of the pins is used to form a self image. The same procedure is followed except that imaging lenses are no longer required. We show that there is no advantage in using self imaging with this approach.

In all current Fourier transform processing systems, which we call conventional Fourier transform (CFT) processors, no matter what kind of filter is used, its filter function can be expressed as a diagonal matrix, if in the view of digital image processing. We have presented a generalized Fourier transform (GFT) processor by extending the diagonal filter matrix into a nondiagonal matrix. It includes CFT as a special case, and still retains the space/time- invariance property. In this paper, we present a method based on genetic algorithms for finding an optimal filter of GFT processor. The behavior of the optimal filter in GFT processor and its advantages over that in the CFT processor are illustrated by the satisfied test results. An optimal generalized Teoplitz matrix for the GFT processor filter based on the figure of merit--the Manhatten error norm is also proposed.

We design an incoherent correlator based on the shadow casting principle, including a bipolar representation of correlation filters. This correlator is totally new to our knowledge, and is furthermore very simple and low cost, although it can not handle large resolution images. The bipolar technique allows us to represent any linear filter in the correlator. We demonstrate experimentally its efficiency in the case of optimal trade-off filters.

The chirp encoded joint transform correlator (JTC) with separate input planes has been demonstrated in two experiments using Michaelson and Mach-Zehnder interferometers to combine the inputs. In the first, the Fourier plane modulation characteristics of the chirp- encoded JTC was investigated by correlating two pinholes. The peak-to-noise ratio performance in this experiment was 26 dB. In the second, chirp modulation was applied to correlating one fingerprint against a set of four fingerprints. Binary ferroelectric SLMs were used as inputs and for display in the Fourier plane. A discrimination ratio of 8.1 dB was demonstrated for the chirp modulated JTC. This is a 2.3 dB improvement over the discrimination observed using the same equipment in a binary JTC arrangement.

High level image processing and pattern recognition tasks require multiple correlations. So multichannel correlators are of interest. Some architecture configurations have been presented for the Vander-Lugt architecture, but few have for the JTC. We present the experiments concerning two multichannel JTC architectures and compare some of their performances. One is based on the use of spatial light modulators with partially filled pixels, in the second Dammann gratings were used.

A new geometric viewpoint is presented for optimizing a binary amplitude filter based on finding an ordered set of phasors, the uncoiled phasor set (UPS), from the filter object's discrete Fourier transform that determines a convex polygon. The maximum distance across the polygon determines the value of the correlation peak and the set of frequencies that the optimal filter should pass. Algorithms are presented for finding the UPS and the maximum distance across the polygon that are competititve with optimization approaches that use the binning (Farn and Goodman). The new viewpoint provides a simple way to establish a bound on binning error.

A real-time optical processing system with dual liquid-crystal spatial light modulators is presented that have been used as an amplitude input and a multilevel phase-only filter, respectively. Numerical calculations are performed for a gray-level and binarized amplitude- phase correlations. An improvement of performance criteria such as discrimination capability, light efficiency, and signal-to-noise ratio has been made for an amplitude in a binary mode to phase correlator. The higher the threshold level of binarized objects is, the better performance criteria produce. Optical experimental results supporting with calculations are described.

A hybrid optoelectronic morphological processor built in the shadow casting architecture is presented. The processor is programmable in both input planes. Binary slices of an input grey scale image and a structuring element of arbitrary size and shape are introduced into the processor by means of Epson liquid crystal spatial light modulators composed of 320 by 264 pixels. We present the quality of optoelectronically made morphological operations in comparison with digital results. The degree of conformity of the results obtained with both methods is measured in terms of the mean (per pixel) absolute error. An analysis of accuracy of the hybrid processor is presented. Several sources of errors, such as for example diffraction effects are discussed.

In designing correlation filters for implementation on cross-coupled spatial light modulators, we are faced with the task of selecting the optimum gain and angle. Usually, this is carried out via an iterative search. In this paper, we introduce a direct method for identifying the optimum gain and angle. Simple simulation examples are included to illustrate the basic idea.

We consider a correlation metric form that is a ratio of quadratics and includes system noise modeled as additive at the correlation plane. We have its optimization on arbitrary spatial light modulators. By selection of coefficients, it may be balanced among a class of individual performance metrics, including synthetic discriminant function criteria. We examine some criteria for selection of the balancing coefficients, so that to one extent or another the resulting filter exhibits robustness against input noise, sharp correlation peaks, invariant response to in- class training images, or discrimination against out-of-class objects.

We previously proposed and implemented a joint transform correlator (JTC) using an optimal trade-off synthetic discriminant function (OT-SDF) filter in order to provide in-plane rotation ivariance. We propose to improve that system by using what we call a synthetic discriminant estimating function (SDEF) filter which also estimates the object rotation angle (without degrading the discrimination capability) through modulating the phase of the correlation peak. Most SDFs/SEFs (synthetic estimating filters) which were real-constrained so far, used the correlation peak height both to determine the object class and to estimate the varying parameter (here the rotation angle); a single correlation could be ambiguous and was not suffiicient. We propose to use the amplitude of the correlation peak only for the discrimination and its phase (which was a free parameter up to now) to indicate the object orientation: we constrain the correlation peak modulus as before, but also constrain the correlation peak phase to be a linear function of the input object rotation angel. Now, discrimination and orientation estimation can be performed simultaneously.

The practical calculation of optical correlation filters in correlators that use spatial light modulators with discrete elements is based on the assumption that the image on the input modulator can be modeled as a modulated 2D comb function or 'bed of nails'. A 2D discrete Fourier transform (DFT) is used to calculate a filter that is also modeled as a modulated bed of nails. The sample values in the comb array are assigned to pixel values in the filter. This approach actually gives fairly good qualitative results in modeling correlation behavior. However, it cannot account in detail for the finite size of pixel elements. The DFT approach has problems when modeling modulators whose pixels' center positions cannot be aligned with corresponding sample values. A modified DFT algorithm and an interpolation scheme for modeling these situations is given. As a practical application of the method, we look at modeling an optical correlator whose pixels are not centered at positions that correspond the DFT sample values.

The design and demonstration of a quantum optical genetic algorithms computer is described. We show that by avoiding the tedious computations of conventional genetic algorithms, time and energy could be saved. The role of quantum indeterminacy as a major component of the operation of this porcessor is emphasized.

Diffraction tomography for medical ultrasound imaging provides increased resolution over conventional pulse-echo ultrasound. This paper describes the generation of ultrasound tomographic images using the Essex ImSyn^TM optical processor. Capabilities include spatial resolutions to as fine as one quarter of the sensing wavelength and a high dynamic range of up to 80 dB. The magnification, field of view, and dynamic range are electronically controllable. High resolution images of an experimental test structure are presented.

In an incoherent imaging system using the liquid crystal active lens, a point spread function can be controlled electrically. Since the liquid crystal active lens is a lens attached to an electrically addressable liquid crystal phase modulator, and the point spread function is an intensity distribution of the Fourier transform of the modulator's phase distribution. In this study, an electro-optical learning system is employed to produce a desirable point spread function by optimizing the phase distribution. Th phase distribution is controlled by a computer and the point spread function of the optical system is captured and fed into the computer. The point spread function is evaluated by the computer to optimize the phase distribution. The optimization method is based on the simulated annealing. The experimentally obtained point spread functions are closer to the desired ones compared to those generated only using a computer, becuase the electro-optical system learns the characteristics of the optical system during the optimization process. The performance of the electro-optical system is discussed.

This paper presents a new image retrieval method suitable for a large-scale image filing system used in, for example, medicine. Keywords for image retieval may be more ambiguous than those for document retrieval, and the meaning of the keywords depends on each user. In order to overcome this problem, we introduce two concept spaces into the retrieval method: one of the user, and the other of the system. The mapping relation between the two concept spaces is then updated through the learning process, so that the retrieval system can be optimized for each user. An optical implementation of this system is also proposed to realize a fast retrieval.

This paper describes the design of an optical sequential machine. The architecture consists of two major components: a combinational logic and a set of memory elements. Polarization- coded symbolic substitution is employed to implement the combinational logic. The memory elements are implemented using a combination of liquid crystal light valves and polarizers.

In most pattern recognition systems, many target patterns should be recognized at one time by fewer reference patterns when the system uses a matched spatial filter. In this paper, two approaches are described for reducing the number of reference patterns in matched spatial filtering called universal pattern recognition which means a recognition technique of various shapes of patterns not depending on a target shapes. One approach is based on very simple bar patterns for the reference object for a matched filter according to a concept that every pattern or symbol is synthesized by many line components with various directions. By also using correlation diagrams with such reference patterns, 26 English alphabets were fairly recognized. Another technique is based on the algorithm to find automatically the unique and proper reference patterns of the matched spatial filter for recognition of the desired target sets. By the algorithm, unique and minimum numbers of reference patterns are selected and 26 English alphabets and 10 digits were recognized.

The problem of crosstalk in the optical resonator neural network based on double correlator architecture is digitally analyzed. Minimum average correlation energy filter (MACE) reducing the crosstalk has been used as a first storing hologram. The comparison of the results obtained for nonorthonormalized pattern by using the MACE filter with those obtained by orthonormalization procedure and phase-only filter is presented.

An algorithm to choose the region of support in a filter whose entries are either 0 or a continuous phase is presented. Such filters will be termed ZAP filters since their entries come from the set consisting of 0 and phases (complex numbers of modulus one). This paper is an immediate sequel to ideas presented previously. As in this previuos work, the algorithm involves a simple thresholding technique applied to the general spatial filter whose design is described. In the technique described the choice of threshold is made by the user in an ad hoc manner. Here the optimal threshold is chosen to be that threshold whose selection results are sometimes surprising. Examples of such are presented here. Having fixed those pixels which are 0 in the ZAP filter, one can then view the remaining phases as free parameters and design a good ZAP filter by optimizing the filter's signal-to-clutter ratio using the author's techniques. One can then optimally discretize this ZAP filter into one whose entries are 0 or one of the nth roots of unity. For example, ternary (o, +1, and -1) filters are of this form as are filters whose entries are 0 and the fifteenth roots of unity. These ideas form the basis for a sequence of computer codes which provide a completely automatic way to start with a training set of true and false targets of any size and automatically and without human intervention produce optimized filters for use in extent and future optical correlators.

In this paper we discuss a new type of time integrative photorefractive device using self- pumped phase conjugation. This device can be used both for demultiplexing and phase sensitive detection.