In this paper we describe some of the research presently being done in areas that will have an impact on the future development of digital optical devices. We will mention: nonlinear optical effects in semi-conductor multiple quantum well material and some device applications; new studies of sub-picosecond nonlinear optical glasses and a figure of merit for material evaluation; and picosecond optical pulse shaping and coding. We conclude by identifying some key directions for future research.

linage rotations translations, changes of scale arid distortions require special pattern recognition and classification techniques. One approach is to look for features such as Circular harmonic components that remain invariant under such changes. Tradeoffs involved in invariant pattern recognition will be discussed, with emphasis on optical implementations.

We discuss the algorithm that enables a machine vision system to recognize a three-dimensional object from a series of multiple views. The algorithm uses pseudo reflection tomography to map N images of the object into a single signature image that incorporates information from all N views. Rotation, scale, and brightness-invariant features are extracted from the signature image that enable robust recognition of the object.

Scale- and rotation-invariance in pattern recognition by optical spatial filtering procedures is discussed using general considerations. As a result a novel approach is proposed and demonstrated by an example employing bipolar spatial filters. The work is concluded by presenting some preliminary results obtained with computer simulation.

Irradiance moments are very adequat for object recognition, since they give an approximation of increasing accuracy and object characteristics independent of shift, rotation and scale can be calculated thereof. Irradiance moments can be obtained by optical detection of the derivatives at the origin of the Fourier spectrum. The quantitative determination of these moments from measured intensity and phase samples in the optical Fourier spectrum has been studied theoretically and experimentally. From intensity measurements only second and forth order moments can be deduced. First and third order moments were obtained from the phase, measured in a 12 point array using heterodyne detection.

A Monte Carlo method based on simmulated annealing is investigated for phase retrieval. Its performance is found successful for reconstructing good estimates of objects with high contrast areas.

The Optimal Correlation Filter for the discrimination or classification of multi-class stochastic images buried in additive noise is designed. We consider noise in images as the (K+1)th class of stochastic image so that the K-class with noise problem becomes a problem of (K+1)-classes: K-class without noise plus the (K+1)th class of noise. Experimental verifications with both low frequency background noise and high fre-quency shot noise show that the new filter design is reliable.

New architectures are discussed for the implementation of signal transforms using new techniques in optical signal processing and computing. These architectures are based upon the factorization of any discrete trigonometric transform matrix into a simple preprocessing matrix followed by a matrix consisting of direct sum of circular convolutions.

Nonlinear optics can contribute decisions to optical signal processing and computing. Etalons permit massive parallelism and global interconnectivity. GaAs etalons appear more attractive for real systems, but ZnS interference filters are convenient for simple demonstrations of logic operations, pattern recognition, and symbolic substitution.

A four level system such as found in dyes and F2+ centres of alkali halides is irradiated by two fields in a Fabry-Perot or ring cavity.It is shown that both fields exhibit optical hysteresis and bistability.Applications such as a single element half-adder and tunable optical bistability are described.

Subnanosecond single-beam thresholding and a high-contrast, high-gain optical AND gate are demonstrated using a nonlinear Fabry-Perot InGaAsP/InP laser amplifier. These devices switch on or off in less than a nanosecond, dissipating only femtojoules of optical input energy. They have large gain (10-20dB) and are operated using diode laser sources. The switching dynamics are theoretically modeled. Potential applications of arrays of these devices for parallel signal processing and computing are discussed.

The effect of external noise of different spectral distributions on the dynamic behaviour of an intrinsic optical bistable device is studied. Our experiments are performed under absorptive conditions in a sodium-filled nonlinear Fabry-Perot in the bad cavity limit. Alternatively, a bistable electronic circuit (Schmitt trigger) is shown to exhibit the same type of noise phenomena while admitting more detailed investigations.

The non-linear wave-equation has been solved numerically for a Fabry-Perot etalon, assuming a cylindrical symmetric gaussian input beam, a local mixed dispersive and absorptive saturable nonlinearity and an additional nonsaturable background absorption term. The role of focusing has been systematically investigated for both absorptive and dispersive bistability. Whole-beam switching occurs under tight focusing conditions. But, in the case of defocusing and absorptive bistability, a further reduction of spotsize results in loss of the hysteresis characteristics. Thus, thisposes a restriction on the minimum spotsize and, hence on the minimum input power, for bistable operation.

A recently presented eikonal treatment of nonlinear resonators and coupled resonators is generalized to enable the treatment of absorptive as well as dispersive optical bistability. The behavior of molecular media, incorporated in various types of resonators is emphasized.

An experimental study is presented of dynamic instabilities in optical bistability with two-state atoms in a cavity. The basic conceptual simplicity of the system as well as the degree to which it has been quantitatively characterized allows precise comparisons between theory and experiment. Instability boundaries in the parameter space of atomic and cavity detunings, of atomic cooperativity and of intracavity intensity are discussed and compared with theory.

A Bimodal Optical Computer (BOC) is presented. The speed and accuracy of the BOC are studied and compared with that of the digital computer. Accuracies comparable to that of the digital computer are achievable, while BOC speed is far more superior in solving a system of linear equations. This advantage in speed increases with the increase of the size of the matrix. The problem of the convergence of the solution using the BOC is investigated. It is found that even with using electro-optical systems with an error as high as 50% in the optical mask and 1% in the electro-optical devices, convergence was achieved for matrices with condition numbers of 25. The effect of the condition number on the convergence of the solution is investigated. It is found that matrices with large condition numbers converge very slowly. Convergence for matrices with condition numbers of higher than 250 was achieved. Some means of improving the condition number of matrix are also introduced.

Real time texture processing is hinded by the large amount of computation required by the texture measurement. Laws' digital method of texture analysis presents the double advantage of being particularly discriminant and perfectly suitable for an optical preprocessing operation based on birefringent crystal properties. Associated to an incoherent to coherent converter, such an optical preprocessor could yield a quasi real-time high-resolution measure of texture, the texture classification and segmentation being done by a digital computer. We shall also present results showing the feasibility and the resolution that may be achieved by such a texture processor.

In Fourier optics, the Fourier transformation is performed on the amplitude of the electromagnetic wave, whereas in the computational method the transformation is usually applied to the grayness of its image which is proportional to the intensity of the electromagnetic signal.

A review of spatial light modulator technologies is presented, including the decription and performance parameters of the main devices, general issues, and their main applications for optical information processing and displays; emerging new technologies as well as future prospects for spatial light modulators are discussed.

We describe a computer-optical processing system that uses an inexpensive liquid crystal (LCD) television monitor and a selective holographic filter for coherent pattern recognition. Specifically, we use a digital computer to generate an edge enhanced image of an object, expose a Fourier transform hologram of this image, and use the hologram as a sort of matched filter for recognizing the original object in real time.

An optical matrix processing system based on an edge-addressed electro-optic modulator can perform matrix-matrix multiplications, correlations, convolutions, cross ambiguity functions, and matrix inversions. Architectures have been developed for handling bipolar and complex data, as well as for the compensation of square-law detection nonlinearities.

A new way to modulate coherent light is achieved by using acousto-optic and electro-optic diffraction simultaneously in surface acoustic wave devices. The light is contained to a region directly under the electro-optic transducer. Simultaneous use of electro-optic and acousto-optic interactions causes the interaction between them that can be used to give multiplication between the acousto-optic and electro-optic signals. The object of this paper is to fully describe this interaction and also show how it can give more degrees of freedom to the designer and lead to more versatile and compact implementation of optical computing architecture.

The concept of a two-dimensional space to temporal-frequency converter was described previously ^1,2. Some technological considerations and system approach were presented in ^3,4 . In the present manuscript the utilization of the converter as "real time" image processor will be shown.

Imaging through a dynamic scattering medium produces low contrast image due to additional large background of scattered light. A passive phase conjugate mirror (PPCM) can generate an image with enhanced contrast. The background noise changes faster than the recording medium integration time and therefore is not recorded. This system does not require a reference wave coherent with the image and it operates in real time. Experimental results through a simulated scattering medium demonstrate contrast improvement over that obtained with conventional imaging arrangement.

We describe various image bearing oscillations originating in one, or two coupled photo-refractive crystals, which can support phase and amplitude modulation as well as multimode fibers. This includes a first report to our knowledge of the operation of the Double Phase Conjugate Mirror. This four-wave mixing device can be pumped simultaneously by image bearing beams originating from the same or different lasers, and couples these two beams into each other. These configurations have potential applications in real-time image processing.

After a brief overview on basic analog optical operations we present image amplifiers being realized by two- and four-wave mixing in photorefractive BaTiO3. We summarize some optical operations with photorefractive elements and present for the first time an image transmitting optical feedback system with gain by a phase-conjugating mirror.

When recording holographic optical elements it is often necessary to resort to complex wavefronts which can not be obtained with standard optical components. Such wavefronts can be obtained with aspheric elements or computer generated holograms which introduce many difficulties. Instead, we developed recursive design technique for obtaining the complex wavefronts from relatively simple holograms. These techniques are based on changing the geometries and/or the wavelengths between recording and readout of these holograms. We shall present our design techniques with a specific illustration of a holographic Fourier transform lens. We designed the holographic lens with complex wavefronts which were derived from first step hologram. The results demonstrated much lower aberrations and better resolution over a field ranging to ±6°, than comparable spherical holographic lenses.

The two-wave mixing of the recording beams at a photorefractive crystal during real-time holographic recording is used for operating an active system for the stabilization of the holographic setup. A previously recorded permanent hologram does also perform two-wave mixing and may similarly operate the stabilization system. The performance of both stabilization methods are quantitatively characterized and compared.

Optics is attractive for interconnects because the possibility of crossing without any interaction multiple light beams. A crossbar network can be achieved using holographic elements which permit to connect independently all inputs and all outputs. The incorporation of dynamic holographic materials is enticing as this will render the interconnection changeable. However, it is necessary to find first a passive method permitting to achieve beam deflection and secondly a photosensitive material of high optical quality requiring low power levels to optically induce the refractive index changes. We first describe an optical method allowing to produce very large deflections of light beams thus enabling to randomly address any spot on a plane. Such a technique appears applicable to both interconnections of VLSI chips and random access of optical memories. Our scheme for realizing dynamic optical interconnects is based on Bragg diffraction of the beam to steer by a dynamic phase grating which spacing and orientation are changeable in real time. This is achieved in a passive way by acting on the optical frequency of the control beams used to record the dynamic grating. Deflection angles of 15° have been experimentally demonstrated for a 27 nm shift in the control wavelength. For a larger wavelength scanning (50 nm), 28° deflections are anticipated while maintaining the Bragg condition satisfied. We then discuss some issues related to photosensitive materials able to dynamically record the optically induced refractive index change. The specific example of Bi12 Si 020 or Bi12 Ge 020 photorefractive crystals is presented. Indeed these materials are very attractive as they require low driving energy and exhibit a memory effect. This latter property permits to achieve numerous iterations between computing cells before reconfiguration of the interconnect network.

An optical setup which uses an addressable reflective component enabling free space interconnections adaptable to electronic integrated circuits is described. Experimental results indicating a fan-out capability of such configurations are presented.

This presentation deals with the use of an incoherent optical processing device for optical interconnections in digital systems. Our device is based on a lenslet-array/mask assembly. The lenslet array creates multiple images of the input. The mask controls the interconnections pattern by blocking parts of these images. The device can operate with polychromatic and spatially incoherent light, so it can use many electro-optical light sources. Device limitations and capabilities are analyzed. This analysis shows that up to 107 elements may be connected by such devices. A demonstration of a simple digital circuit is presented.

The acousto-optic and electro-optic effects have been used in a wide range of devices for signal processing. A new device, which combines both effects in one integrated manner using surface acoustic wave (SAW) technology, has been implemented and offers several advantages. The objective of this paper is to demonstrate use of this device in optical matrix processing, optical excision, synchronization of spread spectrum signals, and real time ambiguity function generation.

Fiber optic VLSI LANs which use optical processing of the medium access protocol are investigated. The experimental demonstration of all-optical networks utilizing asynchronous code-division multiple access or spread spectrum at 100 Mbps and fixed assignment time-division multiple access at 500 Mbps is reported. Both passive and active star architectures are considered. The active star is implemented with a self-routing LiNbO₃ integrated optic switch. Novel fiber optic interconnect devices required to construct such a network are demonstrated using a laser-assisted etching technique.

The design principles are presented for a new class of homogeneous thin-film waveguide lens. The lenses consist of many lens elements with circular boundaries and are amenable to conventional photolithographic methods of manufacture. The design gives very low spherical aberration for axial input beams with apertures up to f/3. By modifying the lens effective refractive index the focusing performance can approach the diffraction limit for input angles up to 8° from the axis.

We briefly highlight recent CMU (Carnegie Mellon University) research on optical Artificial Intelligence (AI) processors for scene analysis. This work includes new shift-invariant AI correlators, plus symbolic, model-based, associative memory, knowledge-based, relational-graph, and neural optical processor results.

Two ternary, an ordinary ternary (0T) and a binary balanced ternary (BT), number representations to be used for optical computing are discussed. An unsigned OT number is represented by a string of symbols (0,1,2), while for the BT, the three logic symbols take on the set (-1,0,+1). The BT symbols can represent a signed number. Using a particular binary encoding method, the three ternary symbols are converted to a pair of binary symbols. The binary coded ternary (BCT) representation has two advantages. First, it allows the use of the well-developed binary optical components. Second, it reduces the number of input-output chan-nels and thus is able to conserve the optical space-bandwidth product. As an example for arithmetic operations, BCT full addition algorithms are given. As examples for multiple-valued logic computing, BCT Post, Webb, and residue logic elements are discussed. Optical implementations of various BCT arithmetic and logic operations are described. Using the two-port Sagnac Interferametric switches (TPSIS), a number of implementation examples are presented.

Optical logic functions can be obtained through nearly-degenerate four-wave mixing in laser diodes operating above as below threshold. In this nonlinear mixing process the pump frequency w is given either by laser oscillation or by amplification of an external pump signal. The conjugate frequency w + δw is then obtained by intracavity nearly-degenerate four-wave mixing when a probe beam of frequency w - δw is injected through the laser diode colinearly with the pump beam. Several examples using this technique are discussed including high contrast bistable memory, AND, OR, and INV gates.

Information processing systems require the data in coded form. Some generation procedures are described for pulse code, pulse width, and pulse density modulation. In case of sampling the idea of error diffusion is introduced. This concept has been developed to generate the pulse density modulation, but can also be used to advantage in other modulation techniques, e.g. pulse code modulation, to increase flexibility.

We show that logical inferences can be conceived in optics, by the registration of coded diffraction gratings, including a state of undecidability. This approach is a first step towards the optical representation of Artificial Intelligence.

In a recent paper, Eichmann and Caulfield) presented a preliminary exposition of optical learning machines suited for use in expert systems. In this paper, we extend the previous ideas by introducing learning as a means of reinforcement by information gathering and reasoning with uncertainty in a non-Bayesian framework². More specifically, the non-Bayesian approach allows the representation of total ignorance (not knowing) as opposed to assuming equally likely prior distributions.

The optical implementation of computing systems whose structure and function are motivated by natural in-telligence systems is a subject that involves optical computing and neural network models for computation. These are two areas that have individually received attention in yecent years and they share the common property that they promise to provide solutions for fundamental problems in computation. In the case of optical computers the limitation that is being addressed is communication. With optics it is possible to have large arrays of processing elements communicating with each other without connecting a wire between each pair. The need to provide wires for communication in an electronic circuit is perceived as a major technological limitation of VLSI [1]. The primary justification for optical computing is therefore to extend the communication capability in computers [2,3]. It is not clear however precisely how this global communication capability can be put to good use. Through free space inter-connects and volume holograms we can have thousands of computing elements all talking to each other at the same time. Is it possible to do useful computation in such a system? We look at neural network models for an answer to this question.

Storage recipes used in forming associative content addressable memory (ACAM) based on models of neural nets are known to generate false spurious memory states together with those being intentionally stored. Effective use of such memories, and of recent optical embodiments of them, may require either exorcising the spurious states or devising a means for discriminating against them when they are evoked. In this paper an opto-electronic spurious memory discriminating circuit is described. It is used in conjunction with an ACAM, is designed to ignore spurious states when they occur, and indicates which of the nominal states is evoked by the stimulus achieving thereby robust recognition. The latter recognition circuit discussion is used as a vehicle for pointing out distinctive features of the "neural net" approach to signal processing as opposed to other approaches to performing the tasks described.

An extension of Hopfield model of neural networks is proposed, considering sets of eccuivalent versions of a given memory. A logical validation switch is introduced, whose state is determined by the proximity of a probe vector to the respective equivalence classes. The algorithm is discussed in terms of cost and performances.

The principle of information retrieval by association has been suggested as a basis for parallel computing and as the process by which human memory functions.1 Various associative processors have been proposed that use electronic or optical means. Optical schemes,2-7 in particular, those based on holographic principles,^3,6,7 are well suited to associative processing because of their high parallelism and information throughput. Previous workers8 demonstrated that holographically stored images can be recalled by using relatively complicated reference images but did not utilize nonlinear feedback to reduce the large cross talk that results when multiple objects are stored and a partial or distorted input is used for retrieval. These earlier approaches were limited in their ability to reconstruct the output object faithfully from a partial input.

We describe several ways to use photorefractive phase conjugators as thresholding elements for optical associative memories. The first involves background illumination of the crystal of a self-pumped phase conjugator. The second method uses a photorefractive phase conjugator designed so that its reflectivity increases with increasing signal strength. The third method uses semi-self-pumped phase conjugators in which one pump is externally provided as a thresholding reference and the other is self-induced as an oscillation beam. Two examples are the double phase conjugate mirror and a new type of phase conjugator based on injection of a counterpropagating pump beam into a unidirectional ring resonator.

A new implementation of the Hopfield model for neural networks using photoactivated electrically-conducting biological materials is introduced. This implementation allows rapid electrical reprogramming of the network. These neural networks show exciting promise for content addressable memories, image processing, and constrained optimization problems.

Optical digital processing offers an opportunity to go beyond the speed limit of electronic computing mainly because optics. unlike electronics, is inherently suitable for highly parallel architectures. In this presentation we shall not discuss methods of making faster switches; rather, we shall look at how to use incoherent light to allow more gates to work at the same time. In an earlier work' we demonstrated the use of a lenslet array based device to provide interconnections for a simple binary "circuit." Other workers discussed the use of shadow casting for other signal routing problems.2 3 It is also possible to use OTF synthesis to provide convolution-like digital inter-connections, appropriate for problems such as image processing. Particularly, holographic incoherent OTF synthesis appears appealing because of its inherent flexibility. Here, we shall consider some candidates for future research in digital information processing with incoherent light. Some of these are Non-binary ("multiple value") digital logic;Cellular architectures; Neural model associative memory.

A special algorithm for the Hadamard transformation (HT) is proposed. It decomposes the space-variant HT into a sequence of convolution and undersampling operations, thereby simplifying the interconnection scheme for the transformation. This allows parallel processing of high-bandwidth signals, either 1-D time signals or 2-D images. We discuss two possibilities for the optical implementation of the HT.

A Space Integrating (SI) Optical Linear Algebra Processor (OLAP) is described and laboratory results on its performance in several practical engineering problems are presented. The applications include its use in the solution of a nonlinear matrix equation for optimal control and a parabolic Partial Differential Equation (PDE), the transient diffusion equation with two spatial variables. Frequency-multiplexed, analog and high accuracy non-base-two data encoding are used and discussed. A multi-processor OLAP architecture is described and partitioning and data flow issues are addressed.

Experimental results of conoscopic holography are presented in this paper. Conoscopic holography is a new method for forming holograms using incoherent light that we recently described. The theory is also expanded to include off-axis holograms.

Single frame close-range photogrammetry was combined with white light projection moire for analysing mechanical properties of a cylindrical shell segment (css). Computer processing methods were developed for measuring and calculating the changes in the radius of curvature and strains due to its static loading.

Holographic interferometry in Fourier plane allows 3D displacement measurement from one recording, with fringe localization in Fourier plane and with the interpretation of only two type of fringe patterns. The measuring range can be controlled by the parameters of the Fourier lens and covers units to hundreds micrometers. Such a Fourier holographic processor with real-time recording is very suitable for interfacing to a computer image processor which works for automatic interpretation of the holographic interferograms and precise measurement.