Previously, we have designed 3-level filters(suitable for implementation on magneto-optic spatial light modulators) to maximize the output signal-to-noise ratio (SNR) and to separately maximize peak-to-correlation energy (PCE) that measures the correlation peak sharpness. In practice, we want the correlation peaks to be sharp (i.e., large PCE) as well as noise-tolerant (i.e., large SNR). In this paper, we present a new method to optimally combine these two desirable properties into a single optimization procedure. Similar methods to trade off SNR versus peak efficiency and PCE versus peak efficiency will be presented. Both simulation and experimental results will be included.

A general algorithm for synthesizing purely real correlation filters in the frequency domain is developed using the method of Lagrange multipliers. This method can be applied to filters which can be derived using linearly constrained quadratic minimization. Purely real minimum average correlation energy (MACE) filters are synthesized to illustrate our approach. Their performance is found to be comparable to original complex MACE filters. The main advantage of our approach is the generation of purely real filters which can be easily implemented in spatial light modulators without holograms and yield the correlation output on the zero-order beam.

We present an automatic method of designing correlation filters for pattern recognition that are composed of select local features (i.e., small parts of a reference object). The local features are selected for their ability to discriminate between the reference object and other known objects or patterns. In the basic localized feature selection problem, we design a correlation filter from a single optimal local feature. In the general localized feature selection problem, we design a correlation filter composed of several local features. We show that the discrimination ability of a correlation filter designed form properly selected local features is actually greater than the discrimination ability of a traditional matched filter.

Correlation s often used as an approach to automated pattern recognition. Generally, correlation provides a measure of the similarity between a reference template and regions of an input image. This measure is also highly dependent on intensity variations in the input image, thereby hindering the performance of simple peak detection decision algorithms. Normalization can be used to achieve intensity invariance of correlation results. This paper addresses some aspects of normalization for a few filter types. For matched filters, the Cauchy-Schwarz inequality provides an effective method by taking into account the energy of the input image within the spatial region of support of the template. For many other types of filters being considered for pattern recognition applications, the regions of support are not always limited to the area occupied by the template pattern. This excessive support can produce undesirable effects in the correlation results whether normalized or not. Benefits of normalized correlation such as intensity invariance and resistance to high energy clutter are discussed along with some problems associated with regions of support. Matched filters, phase-only filters, and binary phase-only filters are investigated. Computer simulations of several cases is used to demonstrate results.

Texture plays an important role in vision research and in many practical applications. Texture research has been wide in applied sciences and many methods for texture analysis have been proposed. To recognize texture a real-time modern optics may offer new tools, which is the theme of the present paper. We investigate the applicability of complex matched filter and phase-only filters to recognize textures. Natural textures of the Brodatz collection were used as the test samples. We show that optical correlators are useful in texture recognition. The recognition is based on the peak-to-correlation ratio, which is much higher for phase-only filters than for matched filters. However, if we consider output signal-to-noise ratio, the matched filter remains valuable.

A first phase demonstration of the capabilities and limitations of an optical correlator in a realistic environment has been completed. The testing was divided into several areas, from laboratory data gathering to a fully functional helicopter-delivered demonstration airframe. The basic research performed has led to three fully fieldable test units which have proven to be rugged and dependable under normal test range conditions. The units were transportable and required no realignment of the optics. Two of the test systems were modular in construction while the third was a 'solid optic' design having optical paths and components contained within a solid glass construction. Two flights have been completed so far, and in both cases the target was identified and tracked, and an airframe guided to target impact.

The ternary phase-amplitude filter (TPAF) is by definition restricted to the modulation values -1, 0, and 1, thus composing a binary phase-only filter (BPOF) multiplied by a binary- amplitude pattern, i.e., a region of support. The TPAF offers an attractive combination of real-time implementation with available devices and good correlation performance. Smart (optimized distortion-invariant) TPAF formulations have been developed. In applying smart filters, mixed optimizations are needed to address not only SNR (signal-to-noise) performance but other metrics addressing practical performance factors such as correlation efficiency. We have developed a design methodology that facilitates mixed optimizations and will report its application to two filter design cases involving a rotated and scaled target on several backgrounds.

A specially designed circular harmonic filter (CHF) which yields bipolar inner product between the filter function and the input pattern is presented. The CH covariance filter which combines the object function and a uniform circular harmonic background is capable of improving the discrimination capability of a single CH filter.

A photorefractive optical correlator capable of recognizing objects regardless of their position, rotation, or scaling in Cartesian space is described and demonstrated. This work is a combination of two well-established ideas. photorefractive correlation is realized using degenerate four-wave mixing in bismuth silicon oxide (BSO). position, in-plane rotation, and scale invariance are obtained using a Fourier plane coordinate transformation. Investigation of the sensitivity to rotation of an image in Cartesian coordinates is reported. In comparison, effective correlation results demonstrating the invariance of the coordinate transformation representation are provided.

Miniature optical correlators provide a means of portable real-time pattern recognition. Most of the attention has been given to reducing length with little consideration to system robustness. Although it is recognized that aberrations will limit system performance, more critical limitations are alignment and sensitivity considerations. This paper presents a procedure for the design of a short robust optical correlator. Theoretical limits on system miniaturization are restricted to component availability; however, practical limitations are dictated by physical constraints and alignment tolerances. This paper explores engineering trade-offs among length, alignability, aberrations, and system reliability. In addition, a complete system analysis including performance, aberrations, and sensitivity is presented.

This paper discusses quadratic filters for distortion invariant pattern recognition. Quadratic filters are easily implemented in coherent optical correlators and are shown to improve clutter rejection. Issues relevant to the design and performance of these filters are discussed and an optimal filter design is derived.

Vision processing is one of the most computationally intensive tasks required of an autonomous robot. The data flow from a single typical imaging sensor is roughly 60 Mbits/sec, which can easily overload current on-board processors. Optical correlator-based processing can be used to perform many of the functions required of a general robotic vision system, such as object recognition, tracking, and orientation determination, and can perform these functions fast enough to keep pace with the incoming sensor data. We describe a hybrid digital electronic/analog optical robotic vision processing system developed at Ames Research Center to test concepts and algorithms for autonomous construction, inspection, and maintenance of space-based habitats. We discuss the system architecture design and implementation, its performance characteristics, and our future plans. In particular, we compare the performance of the system to a more conventional all digital electronic system developed concurrently. The hybrid system consistently outperforms the digital electronic one in both speed and robustness.

Optical pattern recognition techniques for characterizing spherical and nonspherical aerosols and particulates utilized by Dugway Proving Ground in their smoke and obscurant testing programs are presented. The three standard classes of unclassified military smokes and obscurants are included in this study, namely standard liquid aerosol smokes, fibrous obscurants, and flaked brass and graphite IR obscurants. The overall objectives of size distribution in nearly real time under conditions of high particulate loading are specified. A holographic version of the commercially available wedge-ring detector is described to replace the classical matched and phase-only filters, and some of the possible benefits are enumerated for characterizing the two classes of nonspherical obscurants. Computer simulations and experimental data collected for opaque spherical and nonspherical particles are included utilizing the classical matched filter, the phase-only filter and the holographic wedge-ring detector. Computer simulations are presented predicting the performance of a system where the output from the optical correlator utilizing the holographic wedge-ring detector is coupled directly into a optical processor to accomplish the decision making and classification tasks to enhance the speed and performance of the system

This paper examines Bragg diffraction in optical memories and optical correlators which use volume holography. The analysis accommodates the actual structure of the volume grating and does not extrapolate results obtained from plane grating structures. The effects of misalignment, Bragg degeneracy, and cross talk are examined.

In this paper, a hybrid pattern recognition system is described that is based on a matched spatial filter (MSF) synthesized by feature-extracted reference patterns. Some features are extracted from test objects by computer simulations to find the optimum reference patterns for the filter. Through more than 30 simulations, four best patterns are chosen for recognition of 10 kinds of digits by the MSF. In the system, recognition is performed not by conventional autocorrelation peaks but by cross correlation signals because reference patterns for the filter are not exactly the same as the input unknown. The proposed system shows good feasibility for feature extracted pattern recognition.

The integration of optical/optoelectronic processing functions for operating on multiple C3I signals/data including surveillance, electronic support measures, communication, intelligence, and imaging is addressed. The C3I optical processor is capable of operating in a broad spectrum of signals (HF to IR) provided by multispectral passive sensors, bistatic ESM, IR cameras, and multimode/band radars. The processor meets processing requirements for multifunction airborne surveillance and advanced space sensor systems. Attention is also given to optical processors for adaptive null steering, adaptive beamforming, and general purpose computation.

A major drawback of optical correlator systems is the poor quality of correlation signals. Background noise is one of the many sources of false correlation peaks. However, by reconstructing an input scene with the object's amplitude function, an object-enhanced scene - used as input to the optical correlators - (with respect to background) can be obtained. Thus, the optical correlation of this enhanced image improves the signal-to-noise ratio in the correlation image. Furthermore, construction of a filter plane mask that permits simultaneous scene enhancement and cross-correlation operations is described. Experimental results demonstrate the practicality of this approach.

In order to search for symbolically encoded sequences of DNA base information, we have constructed an incoherent optical feature extraction system. This approach uses video display, spatial light modulation, and detection components in conjunction with microlenslet replicating optics, to expedite the recognition of symbol sequences based on their symmetry properties. Multichannel operation is achieved through the replication of input scenery, making possible a higher throughput rate than for single channel systems. A notable feature of our arrangement has been the exchanged positions of input scenery and the filter set. The conventional treatment has been to display the input scene on a monitor for projection onto a set of feature extraction vectors realized as amplitude modulated LCTV devices or lithographically prepared masks. We have chosen instead to provide the filter set as input to the system and to correspondingly place the sequence data in the filter plane of the system, relying on the commutativity of projection to allow this role reversal. A class of DNA sequences known as palindromes are known to have special regulatory functions in biological systems; this class is distinguished by the antisymmetric arrangement of bases in palindromic sequences. We have designed our optical feature extractor to classify short (6 bases in length) sequences of DNA as palindrome or nonpalindrome. We note that this classification is made on the basis of the sequence symmetry, independent of base composition. We discuss the design of this architecture and the considerations that led us to the sequence representation. Initial results of this work are presented. Finally, the integration of this optical architecture into a complete system is discussed.

The problem to recognizing a uniform color object placed on a color background is studied. The recognition process, based on a multichannel correlation, is divided into two steps. First, the presence of the target is determined by using two phase only filters (POF) in each channel matched to the target on a black background, and to a color square with the target in black. The background color is obtained with a classical matched filter (CMF) matched to the latter signal. From the information provided by the three filters in each channel, it is possible to obtain the presence of the target and the character-background color combination. Some simulation results are given to show the performance of the method.

A hybrid pattern recognition system which consists of photorefractive crystal, SLM, and computers is designed and analyzed. In this system, photorefractive crystals are used as dynamic real-time recording materials and optical coherent spatial filters. The problems of excessive sensitivity to scale and rotation in general optical matched filters are solved by computer-aided digital pre-processing. The limitation of angular deviation apart from the Bragg condition due to volume hologram leads to no shift invariant. Thus, the optical properties such as hologram recording speed and diffraction efficiency of FWM in photorefractive crystals are different and strongly depended on the operation condition. We design the most suitable architectures to fit the purpose of the system. Finally, the performance of the whole system is discussed.

We investigate the performance of the binary joint transform correlator (JTC) in the presence of multiple objects for two types of thresholding that are used to binarize the joint power spectrum. The first method uses the median of the joint power spectrum of the reference image and the input scene as the threshold value. The second method is a two dimensional thresholding technique used to maximize the light intensity of the correlation peak and it eliminates the even order harmonic terms. The correlation performance of the binary JTC is determined for both thresholding methods. The binary JTC output is determined analytically in terms of the multiple input targets. The separation requirements of the binary JTC and the conventional JTC in the presence of multiple targets are computed. Computer simulation and experiments are presented for a limited number of multiple target images to determine the correlation peak to sidelobe ratio, and correlation width for both thresholding techniques. In the experiments, a hybrid optical processor with an optically addressed SLM is used to implement the binary JTC. The results indicate that using both thresholding methods, the binary JTC produces large peak to sidelobe ratio and narrow peak for the multiple targets images used in the tests. The two dimensional threshold function produces better correlation performance compared with the median thresholding.

We describe an optical fingerprint identification system that optically reads a latent fingerprint for correlation using a binary joint transform correlator. The fingerprint is read using the total internal reflection property of a prism. The system was built and tested, and the experimental results are presented.

Planar integration of diffractive-reflective optical systems around a glass substrate is an important concept for efficient, reliable, and compact packaging of optical interconnects. Realization of joint transform correlators in planar integrated packages is described. Such compact correlator packages can be combined with other integrated packages on an optoelectronic breadboard to build complex optical processors for various target recognition applications.

An all optical joint transform correlator using a nonlinear optically addressed ferroelectric liquid crystal spatial light modulator (OASLM) is presented. A modification of the nonlinearity characteristic is obtained by simply changing such driving parameters as the amplitude or duration of the voltage applied to the OASLM. The role of those parameters in implementing optically the nonlinear step of the correlator has been investigated here.

A rotation and scale invariant joint transform correlator is described. The polar-logarithmic coordinate transformation is used as a preprocessing of input images. The resultant coordinate transformed images are interfaced to a joint transform correlator. The use of optically addressed ferroelectric liquid crystal spatial light modulators at the coordinate transform plane and at the Fourier plane allows data to be processed faster than video rate and gives the characteristics of a binary joint transform correlator. The rotation and scale invariance of the correlator is demonstrated experimentally.

The quantization effects of the joint power spectrum on the performance of the binary joint transform correlator (JTC) is investigated. Analysis of the quantization effects of the binarized joint power spectrum is presented. The signal-to-noise ratio (SNR) of the binarized joint power spectrum is determined for various levels of the Fourier plane quantization of the binary JTC. The SNR is computed for the truncated binarized joint power spectrum plane and the binary JTC performance is determined for various levels of Fourier plane truncation. It is shown that a capture effect exists in the performance of the binary JTC in the presence of the Fourier plane quantization. Nonlinear compression or truncation in the Fourier plane can improve the performance of the binary JTC at low quantization levels. Computer simulation is used to investigate the sensitivity of the binary JTC and the classic JTC in the presence of Fourier plane quantization. The results of the theoretical analysis are verified by the computer simulations.

A simple method for optical implementation of a composite filter in a joint transform correlator is presented. A composite filter in object space is a real function, but by adding an appropriate constant value, only positive values are obtained. As a drawback, the bias introduced decreases the relative height of the correlation peaks in the recognition process. Non-linear effects in a joint transform correlator are used to increase discrimination by using positive composite filters. Digital simulation and optical implementation of the process has been carried out, achieving the non-linearities by means of saturation effects on the intensity detection process.

A compact, FT lens binary joint transform correlator (BJTC) based on a high resolution ferroelectric liquid crystal, optically addressed spatial light modulator (FLC-OASLM) read in transmission mode is proposed. Experimental results of its implementation using two semiconductor lasers are presented. A binary and a gray scale driving operation of the SLM are introduced and good multiple object recognition capability at high frame rates is demonstrated.

In speckle correlation technique, both the specklegram and its Fourier transformed fringe pattern are detected by nonlinear Optic RAM detectors. The speckle correlation is calculated by the nonlinear optical correlator using a phase modulation property of a liquid crystal television. The effects of the binarization for the signals on the diffraction efficiency are discussed theoretically and experimentally.

Many problems related to optical pattern recognition nay be reduced to the detection and localization of search target. Various modifications of classical matched filte9 such as phase-only filters (POF) binary hase-only filters (Bi-POF) , and complex ternary matched filters (CTMF) providing the output peak sharpness minimalizaton are recently suggested in order to improve the detection quality. Among various matched filtering modifications leading5 to output peak sharpness minimalzation the inverse filtering (IF) , pure phase—only correlation (POC) , hard clipping nonlinear joint transform correlator (NJTC) and8 in some approximation , the optimum filter (OF) proposed by Jaroslawski provide the maximum sharp output signal (which is the diffraction limited delta function) . All procedures introduced to improve the detection quality have high pass characteristics and the high spatial frequencies are enhanced. Most of mentioned above methods (exept POC and NJTC) introduce only modification of the filter function keeping nonchanged spectrum of the input scene. The pure phase—only correlation method belongs to the class of nonlinear matched filtering such that both the filter transfer function and the input scene spectrum are modifed nonlinearly before they are multiplied in the spectral domn9 .Symmetric phase-only filter (SPOF) analyzed by Ersoy et all.9' as an example of the nonlinear matched filtering corresponds exactly to the POC method. Optical implementation of the POC method can be realized on the basis of the nonlinear joint transform correlator7. In this paper analysis of the properties of the phase-only correlation method applied p he multiobject input scenes is presented. Our initial results 1, obtained by digital simulation show that also in this case the method works well, but is sensitive to intermodulation noises. The study of the influence of the intermodulation noises on the recognition results is the main goal of our digital analysis.

We address the problem of optical phase errors in an optical correlator introduced by the input and filter plane spatial light modulators. Specifically, we study a laboratory correlator with magnetooptic spatial light modulator (MOSLM) devices. We measure and characterize the phase errors, analyze their effects on the correlation process, and discuss a means of correction through a design modification of the binary phase-only optical filter function. The phase correction technique is found to produce correlation results close to those of an error-free correlator.

Multiplexed binary phase-only filters provide simultaneously multiple correlations for an input object. A theoretical harmonic analysis allows estimation of the crosstalk among component filters, caused by the nonlinear binarization. A method thresholding a combination of the component phase-only filters is proposed to recover the correlation peak intensities of the component filters. Experimental results using multiplexed binary phase-only circular harmonic filters for rotation invariant pattern recognition is presented.

A method of encoding intensity images as phase-only images will be presented. This method is suitable for implementation in a variety of devices to be used as the input to an optical correlator. Continuous and discrete phase-only filters can be designed to recognize a variety of target aspects of such encoded images. Moreover, there appears to be a distinct advantage in using such an encoding scheme for filters designed to give a fairly stable response over quite large training sets. This advantage appears to be more than four- to-one when compared to filters designed to recognize the standard amplitude inputs. The advantage is particularly pronounced when the phase-only filters are designed to given an optimal response over the detector in the correlation plane. Complete results of computer simulations are given for realistic training sets containing up to ninety-one targets. The performance of the phase-only filters against targets not in the training set is presented, along with results for stability with respect to additive zero mean Gaussian noise.

Binary phase-only synthetic discriminant functions are constructed with a variation of the Jared and Ennis relaxation algorithm and their performance is evaluated on an optical correlator. The correlation responses required in the algorithm can be obtained numerically or experimentally, but the filters constructed by these two different methods are not the same. Reasons for these differences are considered and the usefulness of the resultant filters is investigated.

The construction and principle of operation of a liquid-crystal-over an MOS- array spatial light modulator (SLM) is described. Photographs are presented of patterns produced by amplitude-only modulation of monochromatic light by the SLM when used in a coherently illuminated imaging system. Optical Fourier transforms of phase patterns, which reveal the phase variations, are included. The use of the 16 X 16 pixel array as a pure-phase filter is discussed, and the results of correlation experiments are presented.

An optical correlator using a spatial light modulator (SLM) in the filter plane of half the spatial resolution of an SLM in the input plane is analyzed and evaluated. Reducing the resolution of the filter affected the image used to make the filter by adding a shifted copy of the image to the original. Reducing the resolution of the filter by half resulted in a decrease in storage space and a potential increase in speed by a factor of four. Autocorrelation experiments performed were comparable to those performed at full resolution. In addition, the rotation sensitivity of the filters reduced in resolution were less than that of filters using full resolution.

We demonstrate the use of binary phase-only filters (BPOFs) in a closed loop guidance system for a laboratory lander mockup. Images of a 3-D terrain board taken by the lander's video camera are preprocessed to produce 128 x 128 binary intensity contour maps of the simulated planetary surface. A BPOF is made from a section of the current preprocessed image centered on the exact desired landing site. After the lander has descended to a lower altitude, the BPOF is correlated with a new image. The position of the correlation peak is used in making the next filter and to guide the lander so as to recenter the landing site in the camera's view. We present results of the accuracy with which a site may be tracked from orbit to landing, and the maximum scale, translation, and rotation which can be tolerated between subsequent images. The tolerable scale distortion is quite critical, as it determines the maximum filter update time available at a given descent rate. Application of the results to NASA's proposed Mars Rover Sample Return (MRSR) and Mars Environmental Survey (MESUR) missions is discussed. In both cases, electronic implementations of the algorithm may be sufficient to provide the required guidance system performance.

A genetic algorithm is applied to the task of designing binary phase-only filters in a pattern recognition application. Binary phase-only filters have traditionally been using the classical matched filter as a baseline and then setting the magnitude portion of the filter to unity and binarizing the phase information. The resulting filter has much of its original information content, but is represented with a greatly reduced set of elements. Such filters have been shown to exceed the pattern recognition ability of the classical matched filter on which they are based. However, binary phase-only filters designed using this method are not optimal for discrimination or invariance to pattern changes and several different researchers have investigated various optimization techniques. This paper describes a new technique for designing binary phase-only filters using a genetic algorithm. A population of filters is initially constructed with random phase elements and then modified by the genetic algorithm to produce successively better filters. Each member of the population consists of two chromosomes which contain the genetic information coding for a paid of discrimination filters. During each generation of the algorithm, a new population is produced from the previous population by applying a set of four operators. The four operators include a stochastic remainder selection operator, a two-dimensional crossover operator, a mutation operator, and a survival operator. The fitness function used in the selection and survival operators is based on the ability of the two binary phase-only filters represented by an individual's chromosomes to discriminate between two different classes of characters.

Infrared imagery of 512 X 512 pixels was processed with 128 X 128 arrays by computer simulations of optical correlation. Pyramidal processing using binary phase-only filters was used to process an entire image in parallel at different resolutions. Morphological processing was used to create an image pyramid so that only binary processing was used. The descent policy (processing path to a different level of the pyramid) for optical processing was found to be simpler than that of template matching in digital image processing. Results showed that approximate indications of whether an object is present in a scene can be determined quickly. Further processing generally revealed more precise information. Furthermore, in-class objects present within an image but not exactly matching the filter did not necessarily produce more precise information upon further processing. The termination level of the processing technique is nontrivial and may not be the highest resolution level in some cases.

Correlation provides an effective approach for recognizing targets embedded in a large field of view containing noise and clutter. Correlation is implemented either digitally or optically. When implemented digitally, correlation (consisting ofrepetitive multiply and accumulation operations) is directly realized with digital multipliers and accumulators (MAC), specialized signal processors or general computers. When the input record length is long, digital correlation is implemented in the frequency domain. Fast correlation is achieved because of the existence of the computational efficient fast Fourier transform (FFT) algorithm and specialized hardware for the discrete Fourier transform (DFT).

Implementation of a liquid crystal television (LCTV) for a binary phase only converging beam Vander Lugt correlator system is described. The peculiar pixel array structure of the LCTV characterized by diffraction, scaling, and sampling is presented in a generalized form that can be easily adapted to any pixelated spatial light modulator. It is concluded that twisted nematic LCTV has a potential for becoming an effective and inexpensive spatial light modulator. The twisted nematic LCTV is capable of providing both continuous amplitude and phase modulation and can be used for a variety of real-time optical pattern recognition applications.

Future high performance miniature optical correlators will require specially designed Fourier Lenses. This paper presents optical design considerations and results for several small, high performance Fourier Lenses. These lenses are designed to match up with various spatial light modulators (SLMS) to utilize the full resolution of the SLMS. The miniature designs are compared in size and performance to optics used in the current LMSC Subcompact Hybrid Optical Correlator. The results presented illustrate that the high level of performance of our current correlator can be retained and extended in a miniature correlator, and that the size can be dramatically reduced. The new designs are optimized for use with spatial light modulators having various numbers of pixels and pixel spacings.

By extracting the basic features of an object shape from a digitized Fourier transform image, one can calculate its scale factor and orientation angle. Because of the accuracy of the computed scale and orientation parameters, this formulation can be extended to object classification applications. The concept and system implementation of such a pattern recognition system based on a look-up table interpolation technique for arbitrary-shaped objects are discussed. Finally, the results of performance analysis are presented.

Real-time optical laboratory data is provided on the CMU hybrid optical/digital neural network (NN). Our simulator verifies prior laboratory results and identifies the major system error source as the lack of a true zero input neuron value. Our real-time optical laboratory data includes associative processor (AP), multitarget (MTT) optimization, and adaptive learning NNs.

In order to implement fully adaptive optical multilayer neural networks, a number of issues involving both learning algorithms and device technologies need to be addressed. In this paper, we present some important modifications to existing learning algorithms that serve to simplify optical architectures and allow the use of simple optical devices.

A new architecture for multilayer parallel processing of focal plane array (FPA) images is presented. The RETinally INspired Architecture (RETINA) is inspired by the multilayer parallel structure of biological vision systems. The hardware implementation of RETINA exploits a waveguide hologram (WGH) array generator in conjunction with an integrated spatial light modulator (SLM) array to provide the connection between layers of processing nodes.

Many seemingly diverse optical processing operations and architectures are possible. We unify many of these in one multifunctional inference optical processor for scene analysis.

An associative memory is implemented using a binary phase-only filter as the memory element. In the current architecture, if the input contains any part of the set of stored memories, then the entire set is retrieved at the output. In addition, the sharp autocorrelation peak and the high signal-to-noise ratio allows operation without necessitating a thresholding device.

In this paper, an optical implementation of an associative memory based on parallel rank-one interconnections and time integration is proposed and experimental results are demonstrated. Optical parallel rank-one interconnections based on singular value decomposition and time-integration are useful in implementing real-time, programmable, and adaptable neural networks with arbitrary order. Input patterns can have the same size as the outer-product memory matrix at the expense of time usage proportional to the dimensionality of the input pattern. Computer simulation results are given to show that this time usage can be reduced by using singular value decomposition and principal components analysis, without substantial degradation of the performance of the system. This consequently allows efficient utilization of the spatial parallelism that optical systems can offer.

At Draper Laboratory we are investigating applications of optical processing architectures as they apply to guidance and navigation. In doing so, we have developed a breadboard automatic target recognition (ATR) system that classifies and locates an object regardless of orientation. This system contains a neural network (NN) for object classification and filter selection, and a correlator for extraction of positional information. Correlators suffer from the fact that a given filter is fairly sensitive to variations in scale and rotation of the object under scrutiny. For ATR systems that must locate an object regardless of its orientation, a method of choosing the right filter must be implemented. We have built a system that uses a neural network, implemented in a PC, to classify an object based on the magnitude of its Fourier Transform--a shift invariant function. The result of this classification is then used to select a filter to be displayed on the filter plane spatial light modulator (SLM) in our correlator. The correlator uses a commercial liquid crystal television for gray-scale input and a magneto-optic SLM to display a binary phase-only filter. This architecture not only gives us a non-sequential means of choosing the filter but the resulting correlation can be seen as a confidence factor on the neural network classification. In developing our system we have studied various aspects of NN and correlator performance such as sensitivity to clutter, rotation, and phase aberrations. We have shown this system to be capable of classifying and locating an object regardless of its position and rotation.

Projection method formulation and application for evaluation of the retrieval algorithm and stability condition are presented. By using the theory of the difference equation, we analyze the convergence of the retrieval algorithm in the modified Hopfield content addressable memory (CAM) neural network. A direct analysis of the stability of a trivial solution of the process and the static capacity is also presented.

Two-dimensional acousto-optic (AO) correlators differ from frequency plane correlators in that multiplying, shifting, and adding, rather than Fourier transforming are used to obtain the correlations. Thus, manu of the available composite filter design techniques are not aimed at designing filters for use in AO correlators because they yield frequency-domain functions. In this paper, a method is introduced for designing filter impulse responses of arbitrary extent for implementation on AO correlators. These filters are designed to yield sharp correlation peaks. Simulation results are included to illustrate the viability of the proposed approach. Also included are some initial results from the first successful use of gray-scale composite filters on an AO correlator.

Acousto-optic (AO) cells are widely used in real-time optical processing of RF signals because of their high computational throughput resulting from large modulation bandwidths and large parallel channel capacity. In this paper we discuss how an AO cell can be used to build a general purpose infinite impulse response (IIR) filter for processing sampled data or discrete time signals. The advantages of the system are high speed real time filtering of discrete time RF signals with sampling frequencies in the 10 MHz to a few GHz range. IIR filters are attractive because they can provide a good magnitude characteristics for a smaller number of taps than finite impulse response filters can. In our realization the filter coefficients are stored in spatial light modulators (SLMs) and thus can be changed in time to build adaptive IIR filters for various applications.

A novel optical architecture is presented that is based on the Hartley transform implementation of the classical adaptive least mean square (LMS) algorithm. The Hartley transform is employed to transform the time-domain signal of interest into the frequency domain, where filtering is performed via multiplication by the adaptively controlled filter transfer function. The inverse Hartley transform is then applied to this product resulting in the desired signal estimate, which is then subtracted from the desired signal to obtain the error to be used to update the filter transfer function. In addition to the presentation of this algorithm and the optical implementation, performance assessments based on the number of calculations required for the digital and optical approaches are provided.

An acousto-optic (AO) image correlator architecture will be presented that minimizes the overall system size while maintaining excellent image quality for large input scenes. The correlator can accommodate gray-scale input scenes with dimensions of 512 X 244 pixels and gray-scale reference templates of 64 X 64 pixels. The size of the optical system, however, is less than 10 cubic inches, 1" X 1" X 9". This design incorporates a surface emitting laser diode array that has a center-to-center spacing of the laser elements matched to the row spacing on the CCD. Furthermore, the space-bandwidth and center frequency of the AO cell are chosen to match the length of the input image information in the cell to the width of the CCD. These two design decisions allow close to one-to-one imaging through the entire optical system producing the shortest possible path length. The optics were then designed with a goal of producing nearly diffraction-limited image quality.

A hybrid optical/digital processor has been developed that computes both the magnitude and phase of the bispectrum for wide bandwidth RF signals. The overall optical architecture is that of a modified Mach-Zehnder interferometer which contains three acousto-optic modulators and appropriate transforming lenses. The intensity distribution in the output plane of the interferometer contains an interference term which represents the real part of the bispectrum multiplied by a spatial carrier (the interference fringes). To isolate the bispectrum information, the output image is digitized and digitally filtered. The imaginary part of the bispectrum is obtained by Hilbert transforming the real part, and the bispectrum magnitude and phase are then computed. We demonstrate the performance of the processor with three different test signal sets where the signals have a bandwidth of either 6 MHz or 12 MHz. The test results illustrate the successful recovery of magnitude and phase information for the bispectrum of quadratically related signals.

Optical algebraic processors can perform complex calculations in parallel and at high speeds. However, they commonly suffer from a low analog accuracy which hinders their widespread application. Error detection and correction codes provide one technique for improving the accuracy of optical algebraic processors. The use of these codes would allow some of the errors that may occur during a computation to be detected and possibly corrected. This paper describes the results of various computer simulations of optical matrix-vector multipliers employing error-correction codes. It discusses the application of convolutional codes to optical matrix-vector multipliers along with several block codes. Both binary and nonbinary codes are considered. The results indicate that a significant improvement in performance can be obtained when compared with processors not employing error-correction codes. Also, the type of noise, whether signal-independent or signal-dependent noise, has a significant effect on the performance of a matrix-vector multiplier employing an error code. The encoding and decoding operations required for the error codes can be performed optically.

In this paper, an optical tomographic system based on the Fourier Slice Theorem and using a holographic geometric transformation is described. A new decomposition is first proposed and used for the polar-to-cartesian coordinate transform, which is the most important procedure in the system. The characteristics such as mapping geometry, influence from singular points, the number of pixels required for accurate geometric mapping and system optimization are analyzed and compared with others. The holographic elements and optical systems are designed. The corresponding results in computer simulations and optical experiments, as well as computer-generated hologram (CGH) tests, are provided.

We show that a high-resolution picture of a scene can be generated from multiple low resolution images. The pixels of the multiple images are combined to form a matrix corresponding to the correlation between the required high resolution image and the camera degradation. Examples are shown.

Regularization techniques and Fourier optics are combined to design a class of real time coherent optical processors, which exhibit considerable immunity to phase noise. Regularization and adaptive control is achieved by employing a piezo-electrically controlled displacement of one of the feedback mirrors. The processor has been used to obtain extrapolation of spatially limited objects imaged through a diffraction limited system.

We describe a photorefractive phased array radar processor which directs its look angle to the signal of interest and puts nulls in the array function at the locations of jammers. We derive the array null depth using a simple model of the photorefractive response which includes photorefractive erasure. We analyze the response time of the system, showing that it is independent of the photorefractive time constant. Finally we analyze the response of the system to multiple simultaneous jammers, and to irregular wave fronts establishing that the array response organizes itself to conform to non-planar incident waves. This last property allows the system to be used in complex radar environments and on large irregular nonrigid radar arrays.

Previously, we showed experimental results using spatial light rebroadcasters (SLRs) for orthogonal associative memory and for key word addressable memory. The SLR is suitable for memory because it may be written and then read later. In this paper, a parallel set of half adders is constructed using a single spatial light rebroadcaster together with an image intensifier and optically addressable liquid crystal light valve (LCLV). Experimental results using a 10 X 10 array show 100 additions in one step. We describe how to construct a parallel ripple-carry full adder and a bit-slice full adder with a single half adder.

An optical neural network system using a microchannel plate spatial light modulator (MSLM) and a computer-generated hologram (CGH) displayed on a liquid crystal TV (LCTV) display device are described. In the optical three-layer neural network, the back-propagation learning algorithm is successfully demonstrated with the uniformity of +/- 35% in synaptic connection coefficients in MSLM and the contrast of 20 in neuron output in LCTV, respectively. To make a dynamic optical interconnection and analog optical computing, CGHs are displayed on LCTV. Applications of displaying CGHs, spatial filtering, and moving pictures are performed.

The new high performance liquid crystal over silicon spatial light modulators allows fast electrical modulation, however, such SLMs are binary and pixelated, typically with a small pixel fill factor. This paper investigates the usefulness of these devices as Fourier components, both for optical filtering and as electrically addressed Fourier holographic elements.

An optoelectronic structure for implementing the iterative filtered-back projection image reconstruction algorithm is proposed. The forward projection and back projection operations in the iterative procedure are performed via optoelectronic devices, which include spatial light modulators and charge coupled device detector arrays. The optical Fourier transform is applied to speed up the reconstruction and feed-back is used to minimize the distortion caused by optical transforms.

Two-dimensional spatial light modulator (SLM) structures, employing a novel multiple quantum well (MQW) light valve, are proposed for optical signal processing applications. The optical modulation in the light valve is achieved by the manipulation of field-dependent birefringence in the MQW layers. This is in contrast with the conventional MQW modulators which generally involve amplitude modulation by varying the electroabsorption. Computations of enhanced birefringence (in) due to quantum confined Stark effect (QCSE) as a function of wavelength are presented for the AlGaAs-GaAs MQW layers. MQW light valve configurations using both reflection and transmission mode of operations are discussed. A two-dimensional SLM utilizing InGaAs-GaAs MQW layers, an AlAs-GaAs λ/4 wave stack as dielectric mirror, and an array of integrated InGaAs-GaAs photodetector elements is described.

We study a hybrid optoelectronic architecture for pattern recognition. In this architecture, a multichannel correlator realizes feature extractions on the analyzed image while an electronic neural network (NN) performs the high-level pattern recognition task. Due to its in situ learning and adaptive capabilities, the NN provides an efficient way for full exploitation of the computational power of optical processors. Indeed, not only the theoretical transfer function of the pattern recognition system is realized but also the imperfections of the analog optical computation are learned in the processor. The potential of this approach is illustrated on a simple multiclass problem of robotic classification. precise comparisons with different techniques of filter synthesis for the feature extraction performed by the multichannel correlator are carefully analyzed. An optical implementation based on a joint transform correlator using a photorefractive crystal is presented.

Abstract not available.

Optical algebraic processors can perform complex calculations in parallel and at high speeds. However, they commonly suffer from a low analog accuracy which hinders their widespread application. Error detection and correction codes provide one technique for improving the accuracy of optical algebraic processors. The use of these codes would allow some of the errors that may occur during a computation to be detected and possibly corrected. This paper describes the results of various computer simulations of optical matrix-vector multipliers employing error-correction codes. It discusses the application of convolutional codes to optical matrix-vector multipliers along with several block codes. Both binary and nonbinary codes are not employing error-correction codes. Also, the type of noise, whether signal-independent or signal-dependent, has a significant effect on the performance of a matrix-vector multiplier employing an error code. The encoding and decoding operations required for the error codes can be performed optically.