For practical computer vision applications using optical processors, filters to reduce clutter and detect candidate regions of interest (ROIs) in a scene are necessary before the class of an object can be determined. We discuss new morphological wavelet, Gabor, and rank-order filters to achieve clutter reduction, object detection, and image enhancement. All operations (and hence all low, medium, and high level computer vision operations) can be performed on the same optical correlator architecture with such new filters.

Many correlation filters (e.g., matched spatial filters, phase-only filters, binary phase-only filters, etc.) are usually evaluated in terms of metrics such as signal-to-noise ratio, peak-to- correlation energy and Horner light efficiency. In this paper, we compare the MSF and POF using more direct performance measures, the probability of detection (P_D) and the probability of false alarm (P_FA).

Several methods of designing Synthetic Discriminant Function (SDF) filters exist. All of these require that the correlation output take on specified values at origin. In this paper, we examine the role of these correlation plane constraints. We show that the conventional practice (of forcing the correlation outputs to a constant for all training images from a single class) leads to poor discrimination. We introduce a new method to improve the discrimination capabilities of SDF filters.

We present a new approach to training set selection for general out-of-plane rotation distortion. This approach is mathematically related to distortion-invariant filter design and acts to simplify filter design and improve discrimination performance. It is also applicable to most other model based rotation-invariant filter designs. We develop the approach in a 3-D space with 3-D objects. Given a target object, it is first uniformly rotated about a coordinate axis and sampled. This is denoted as an in- plane rotation set. The set is then rotated in spherical coordinates corresponding in direction to the vertices of a platonic solid. All the rotated objects are correlated to obtain a correlation matrix. The correlation matrix is approximately cyclic Toeplitz dependent on the training set ordering. Its eigenvectors are approximately the Kronecker product of subsets of Fourier Vectors. The training set is intended to be orthographically projected to 2-D images. A filter family is defined by multiplying the training set matrix times the eigenvector matrix. The resulting correlation response at the origin is constant magnitude and linear in phase. As with linear phase coefficient composite filters, we expect such a filter family to demonstrate inherent discrimination against clutter and high signal-to-noise ration in detection.

In general, the output data format of a sensor is not compatible with most programmable spatial light modulators (SLMs) used in optical correlators. Therefore, to use the sensor image as the input image in an optical correlator utilizing a programmable SLM, the image must be transformed or preprocessed. Another reason for preprocessing an image is the inherent edge detection of a binary phase-only filter (BPOF) or a ternary phase-amplitude filter (TPAF). The appropriate preprocessing algorithm can be used to enhance the edges in the input image. To be fully useful in an optical correlator system this preprocessing algorithm should be implementable at throughput rates equal to or greater than the throughput of the sensor. This paper presents the methodology and results of implementing some example preprocessing algorithms in an off-the-shelf image processing board to obtain video frame rates when interfacing to an RS-170 sensor.

Optical correlation algorithms are being developed for implementation in the System for Passive Optical Target Recognition (SPOTR). These algorithms include preprocessing the input imagery for input into the spatial light modulator (SLM), smart filter development for recognizing an M60A2 tank at various elevations, azimuths and ranges, post-processing the output correlation plane to determine its validity, and filter management for efficient implementation of the smart filters.

This paper details a compact two-dimensional optical correlator based on 128 X 128 ferroelectric liquid crystal spatial light modulators in both the image and filter planes. A complete description of this compact correlator is given. The correlator is fully programmable and performs automatic pattern recognition functions at 500 frames per second. Key design parameters and results of performance analysis are presented.

Real-time gray-scale correlation in the spatial domain has been demonstrated previously using an acousto-optical (AO) correlator. This work demonstrates normalized gray-scale correlation as implemented on an AO correlator system capable of operating at real-time video rates. Motivation for using normalized gray-scale correlation is presented. The normalized correlation algorithm as implemented on the AO correlator is detailed. The entire real-time AO correlator system is described, including the electronic support hardware and the user interface. Since normalization requires a division operation, system numerical precision issues are addressed. Test results obtained in non-real time experiments are presented.

The development of an optical correlator system and flight tests to be conducted from a remotely piloted vehicle (RPV) are described. The optical processor is based on laser gyroscope construction techniques and relies on 128 X 128 reflective-mode magneto-optic spatial light modulators for both the input image and spatial filter insertion. The input image is obtained from a visible camera in the nose of the RPV. The processing system incorporates Kalman's invariant filters. The output of the correlator is through a 128 X 128 high speed CCD camera. The correlator system also includes image processing and all electronic drivers. The optical package occupies a volume less than 25 in³ while the whole processor package is less than 1 ft³ and weighs less than 40 lbs, and is ruggedized for temperature, shock, and vibration. The RPV, Eglin Air Force Base test range facilities, tower tests, telemetry, and training set acquisition are discussed.

This paper details a compact correlator unit (CCU) built by Martin Marietta that performs start-to-finish automatic pattern recognition tasks. The CCU is a transportable, programmable two-dimensional optical correlator based on magneto-optic spatial light modulator (MOSLM) technology. A compact transportable correlator presents additional design issues over a conventional benchtop system. This results in a different set of design parameters which are discussed along with analysis and test results of system performance.

A method for designing a general spatial filter as a constrained optimization question is given. This method makes full use of all the free parameters in the design problem. In most instances this method seems to give results which are far superior to those obtained by the synthetic discriminant function technique and its descendants and variants. This is especially true for instances in which there are many true and false targets. The difference between the two techniques may be approximately described as the difference between actually optimizing a rather complicated function and optimizing a linear approximation to the function.

Results of correlations are given using invariant spatial filters made from training sets comprised of Transfer of Optical Processing to Systems (TOPS) images and model-based images. Binary filters were constructed from images with 256 gray levels and from binary images. Correlations with original training set images and with independent data sets were performed. Correlation results as a function of azimuthal and elevation angle extent are presented as well as comparison of real and model image correlation.

Realistic scenery not only has out-of-plane rotation as a distortion but may also contain extreme intra-class intensity variations. We apply a distortion-invariant correlation filter bank known at the Hybrid Composite filter bank to realistic scenery. Hybrid Composite filters are a unification of several well known synthetic discriminant functions which include the classical synthetic discriminant function, Minimum Average Correlation Energy filter, Linear Phase Coefficient Composite filter and Linear Phase Response synthetic discriminant function. it has been demonstrated that Hybrid Composite filter detection and discrimination performance is better than the individual filter types that they are formed from. Hybrid Composite filters act to perform morphological transformations on distorted target and clutter images. The resulting small set of feature shapes can be used to estimate a local peak to sidelobe ratio which is intensity-invariant. Only target training images are used to design the filter bank. Clutter is inherently discriminated against because of the complex phasor relationships used in Hybrid Composite filter design. No scene segmentation is necessary to remove unwanted regions. We demonstrate best case and worst case performance with a set of vehicle images.

Historically synthetic discriminant functions have been designed using specific constraints in the encoding methodologies. In this paper we consider several design alternatives which eliminate the constraints yet appear to perform equally well with the classical versions. These new encoding methods are also found to require a significant reduction in computation with inherently better encoding accuracy. In addition, we discuss a very simple encoding method that allows for a significant increase in the number of references that can be incorporated in a filter.

Adding realistic hardware characteristics to computer simulations of optical processing systems provides insight to the performance of the system. The Photonic Systems Group at Martin Marietta Astronautics has produced an optical correlation simulation library that incorporates the operating characteristics of optical hardware. The simulations using this library modify theoretically perfect simulations for comparison to optical correlator hardware performance. A comparison of this hardware simulation with an actual correlator is presented.

The choice of threshold line angle (TLA) is an important issue in designing Binary Phase-Only Filters (BPOFs). In this paper, we derive expressions that explicitly relate the TLA to correlation peak intensity. We also show some examples that illustrate the effect of choosing the wrong TLA.

Investigations of an algorithm using direct iteration of binary phase states to approach the goal of correlation response uniformity across an arbitrary training set are reported. The approach is applied to a real-world filter design case involving 180 training set images and is compared to other methods (e.g., the Jared-and-Ennis approach) in terms of both effectiveness and computational demands. Concepts to improve and extend the approach and to integrate it with other BPOF (TPAF) optimization techniques that adjust the region of support are suggested.

Ternary Phase-Amplitude Filters (TPAFs) have several advantages in optical pattern recognition applications. These advantages include effective signal-to-noise and signal-to- clutter translation-invariant discrimination ability, relatively high optical efficiency, and implicitly efficient storage of the filters. However, TPAFs are sensitive to changes in scale and rotation. Thus, relatively large numbers of filters must be applied to uniquely identify a particular target image. This paper proposes the use of binary discrimination tree to significantly reduce the number of filter applications required to identify a target image in a real-time target recognition environment.

A method for constructing limited, distortion-invariant, composite, binary filters for an optical nonlinear joint transform correlator is described. Experimental results are presented on the in- situ construction of such filters, their invariance, discrimination, and correlation quality.

We designed binary phase-only filters from a training set of images using a statistical approach. We forced images into clusters and designed filters to recognize objects from that cluster. We report on results obtained by computer simulation comparing the performance of filters to recognize objects from clusters of one and two classes.

This paper is a report on the advanced development and characteristics of a new high resolution, high frame rate, reflected R-MOSLM. This effort is aimed at the production of miniature ruggedized optical correlators (MROC) for optical pattern recognition. Pixel size is under one mil center to center, one third the dimension of present transmission mode devices, thereby reducing the optical path length by an order of magnitude. This development includes optimization of the optical and functional characteristics of the MOSLM for Mil Spec Systems. The device research and process development has been performed at Carnegie Mellon University NSF Data Storage System Center under contract from Litton Data Systems. The Litton Electron Device Division is transitioning the device to production. The MROC system description is described in companion paper (1959-09).

With its constant concern to improve road safety P.S.A. Peugeot Citroen (France) wants to make progress on active safety by developing an autonomous system providing fast practical information to the driver. Recent advances in coupling liquid crystal spatial light modulators (SLMs) with silicon VLSI chips have opened new opportunities in optical parallel processing and computing. It is now possible to construct compact, reliable, and reconfigurable optical hardware for industrial use. We have chosen the task of road sign recognition, with emphasis on speed limit signs, to demonstrate the feasibility of such an optical processor which uses fast switching elements based on ferroelectric liquid crystal technology.

The complex transmittance of an Epson liquid crystal television was determined, as a function of the video drive signal. Several different operating configurations of the LCTVs were established, and their usefulness as operating modes for spatial light modulators in the input- and-filter-planes of a hybrid correlator is discussed. A high-stability phase measurement scheme was developed to determine the operating curves, and this system is presented. The transient response of the liquid crystal television to field-rate changes in the video signal is also presented. It is observed that the switching speed of the liquid crystal molecules presents an obstacle to the operating speed of these devices.

In contrast to imaging and interconnect applications, which require fixed diffractive optical elements, applications such as optical image processing and target recognition require diffractive elements that can be altered dynamically in real-time and, therefore, require the use of spatial light modulators (SLMs). Present SLM technology, however, has limited modulation and space-bandwidth which affects system performance. We present techniques for designing diffractive filters for display on SLMs in coherent and incoherent pattern recognition systems.

We report a computer code that affects several efficiencies in taking a recently developed filter optimization algorithm into practice. The algorithm (Richard D. Juday, `Optimal realizable filters and the minimum Euclidean distance principle,' Applied Optics, in press) analytically optimizes filters on arbitrary modulators (coupled, binary, full complex, fractional 2(pi) phase, etc.). Its quadratic ratio metrics include intensity, signal to noise, and peak to correlation energy. The user supplies necessary information to the program, and also certain common complex arithmetic functions. The code is to be available from COSMIC under the name `MEDOF -- Minimum Euclidean Distance Optical Filters,' NT control number MSC-22380.

The classical matched filter used as the basis for optical correlation filters requires a spatial light modulator (SLM) which can effect fully complex modulation. Practical (single) SLMs cannot achieve this modulation, but many can be made to operate in a `phase-mostly' or `amplitude-mostly' mode. Successful correlators have been constructed using phase-mostly SLMs in the filter plane by disregarding the usually small coupling of amplitude. During the last few years, new theories have been developed on the construction of filters for coupled modulators for a variety of metrics. The liquid crystal cell modulator, which can be operated in a variety of coupled modes, presents a unique opportunity to implement these algorithms. This paper reports on such laboratory measurements.

Among the available filters for pattern recognition, the MACE filter produces the sharpest peak with very small sidelobes. However, when these filters are implemented using practical spatial light modulators (SLMs), because of the constrained nature of the amplitude and phase modulation characteristics of the SLM, the implementation is no longer optimal. The resulting filter response does not produce high accuracy in the recognition of the test images. In this paper, this deterioration in response is overcome by designing constrained MACE filters such that the filter is allowed to have only those values of phase-amplitude combination that can be implemented on a specified SLM. The design is carried out using simulated annealing optimization technique. The algorithm developed and the results obtained on computer simulations of the designed filters are presented.

To our knowledge, the proposed concept is the first non-holographic all optical implementation of an invariant processor. Unlike most of the proposed invariance schemes, our processor works with gray level images, not just binary ones. This follows from the intensity transformation formalized in Eq. (6). Not only the shape of the object but also its intensity distribution can be analyzed for object identification. Nonlinear optical processing is the key element of our scheme.

A real-time pattern recognition system is being developed for use in the adaptive control of welding process. The weld-pool (composed of molten material) is a good indicator of welding quality. Fourier-optical processing is proposed that compares the weld-pool image against templates of various weld-pools shapes of which only one is optimum. If the weld-pool shape most closely matches a particular non-optimal shape, then the weld parameters that cause that aberration can be corrected; e.g., current too high, weld beam off of seam, etc. Correlation of edge-enhanced and windowed images with the template were evaluated. The ratio of matched to mismatched correlations can be improved from 2:1 to 30:1 by removing the electrode from the template and by measuring correlation where the electrode correlation peaks occur. Weld- pool position can be determined from location of correlation peaks. Computer simulations of optical edge detection show that phase-only filter is inappropriate, but magnitude-only filtering yields two edges for each real edge.

Current U.S. Air Force interests include a desire to track an object based on its shape once it has been designated as a target. This paper discusses the use of a correlation based system to track an object through a series of images based on templates derived from previous image frames. The ability to track is extended to sequences which include multiple objects of interest within the field of view. This is accomplished by comparing the height and shape of the template autocorrelation to the peaks in the correlation of the template with the next scene. The result is to identify the region in the next scene which best matches the designated target. In addition to correlation plane postprocessing, an adaptive window is used to determine the template size in order to reduce the effects of correlator walk-off. The image sequences used were taken from a Forward Looking Infrared (FLIR) sensor mounted onboard a DC-3 aircraft. The images contain a T-55 tank and both an M-113 and a TAB-71 armored personnel carrier moving in a columnized formation along a dirt road. The goals of this research were to (1) track targets in the presence of other, and sometimes brighter, targets of similar shape; (2) to maintain small tracking errors; and (3) to reduce the effects of correlator walk-off.

Removing the drift component for terminal descent onto the Martian surface is an important factor for improving the chances of a safe landing. Historically, active radar has been the sensor of choice, but in an effort to reduce power, weight, and volume, we have been investigating the use of an optical processor to analyze descent imagery. Such a processor is a joint transform correlator (JTC) that can compare a current descent image frame with a previously stored image (the reference image). The correlation of overlapping image data provides an output peak that is shifted relative to the autocorrelation peak location at the detection plane. This displacement of the correlation peak from the reference combined with time and altitude information can be processed to provide a drift velocity vector. This paper presents computer simulated results of a JTC processor using Martian terrain board video data. These data were obtained from a camera mounted to a moving base carriage that simulated a realistic descent profile from a scaled altitude of 1,500 m to 100 m.

We describe an architecture of the unidirectional photorefractive ring resonator, in which the competition strengths between spatially confined modes of the ring resonator are externally controllable. We also discuss an application of the ring resonator for an optical high-level constraint relaxation network in the pattern (character or word) recognition system.

We present the recording and reading of color matched spatial filters in photorefractive LiNbO₃ crystals. Detailed analysis and experimental results on the degrading effects encountered are provided as well as techniques to reduce those effects. Topics discussed include color cross talk, crystal thickness effects, scattering, and wavelength selectivity.

A high performance 32 X 32 peak detector array is introduced. This detector consists of a 32 X 32 array of thresholding photo-transistor cells, manufactured with a standard MOSIS digital 2-micron CMOS process. A built-in thresholding function that is able to perform 1024 thresholding operations in parallel strongly distinguishes this chip from available CCD detectors. This high speed detector offers responses from one to 10 milliseconds that is much higher than the commercially available CCD detectors operating at a TV frame rate. The parallel multiple peaks thresholding detection capability makes it particularly suitable for optical correlator and optoelectronically implemented neural networks. The principle of operation, circuit design and the performance characteristics are described. Experimental demonstration of correlation peak detection is also provided. Recently, we have also designed and built an advanced version of a 64 X 64 thresholding photodetector array chip. Experimental investigation of using this chip for pattern recognition is ongoing.

We describe a joint transform correlator in which the reference signal and the input scene are placed in different input planes along the optical axis. We show that this system will focus the off-axis cross-correlation functions and the on-axis autocorrelation functions in different output planes. In a conventional joint transform correlator, these terms are produced in the same output plane. The effect of placing the input scene and the reference signal in different input planes is mathematically analyzed. Computer simulations, and experimental results are presented to illustrate the performance of this correlator. We show that for the nonlinear joint transform correlator, the higher order correlation terms are produced in different output planes.

The performance of the optical joint transform correlator is severely degraded when multiple targets or multiple identical nontarget images in the input plane produce false peaks in the output plane. One method for overcoming this is to displace the input and or reference plane along the optical axis. This results in the true correlation peaks coming to a focus in an output plane separate from the plane where the false peaks come into focus. These planes are also separated by displacement along the optical axis. We show that this separation can be achieved by placing a simple lens system along one side of the optical path in the input side of a conventional joint transform correlator. By appropriate lens choice, any virtual displacement can be achieved. Through use of computer simulation, we show the effects of varying the virtual input displacement.

Experimental results on rapid (2.5 ms, approximately 400 Hz) tracking of a human retina are presented. The optical tracking was implemented with a nonlinear joint transform correlator that used ferroelectric liquid crystal (FLC) spatial light modulators (SLM) at both the input and joint transform planes. The input SLMs were electrically addressed and the joint transform SLM was optically addressed with a photodiode. Realistic images of a human retina were used in a scenario that is directly traceable to surgical applications. For a one hundred image noisy test set, only one false alarm was obtained. The threshold requirements and tracking accuracy were quantified. Hardware improvements that are realizable in the near term (approximately 1 year) are discussed.

An off-axis joint transform optical correlator architecture has been developed which utilizes a novel transmission type photoconductive organic polymer (POP) spatial light modulator. The arrangement allows the writing of `phase only' power spectra within the modulator. The off- axis architecture also allows the separation of the correlation signals from the unavoidable transmission of the writing beams through the modulator. Experimental results are presented during this communication.

Converging light is used for a reference beam in constructing the Fourier transform of a filter for use in a 2-D correlator. This eliminates the need for the inverse Fourier transform lens used for performing correlation. We analyze this approach and demonstrate correct performance with optical experiments. Further, we propose a compact optical correlator system that uses only one curved element for both constructing filters and performing correlation.

A conventional optical pattern recognition system such as the optical correlator is based on template matching between the input target and the filter. However, automatic target recognition, in a heavily camouflaged environment, requires additional spectral information to enable effective target discrimination. In this paper, an innovative spatial-spectral pattern recognition system is introduced. An acousto-optic tunable filter (AOTF) imaging spectrometer is used as the spatial-spectral preprocessor to enhance the camouflaged target image against its background. A cascaded correlator is then used for target recognition and tracking. An overview of the AOTF imaging spectrometer developed at JPL is provided. Experimental demonstrations of camouflaged target recognition from a foliage background are also provided.

An optoelectronic neocognitron neural network has been built at JPL and demonstrated for various automatic target recognition applications. A new technique of constructing and storing a large holographic synapses array at the Fourier plane has enabled the shift-invariant target recognition and tracking. In this paper, a recent development of an innovative bipolar holographic synapses synthesis technique is reported. Experimental demonstration of increased discrimination capability using this bipolar synapses is also provided.

In this paper, a polychromatic neural net using a specially doped LiNbO₃ photorefractive crystal is presented. This neural net is a two-level high-content addressable memory. The first level is a polychromatic Hamming net for color image classification, and the second level is a mapping net, based on a photorefractive crystal associative memory. The major advantage of this neural net must be the large storage capacity and it requires fewer interconnection links. Experimental demonstrations are provided, in which we have shown that the proposed neural net is consistent with the theoretical model.

A complex associative memory model based on a neural network architecture is proposed for recognizing three-dimensional objects acquired from a dynamic environment. The storage representation of the complex associative memory model is based on an efficient amplitude modulated phase-only matched filter. The input to the memory is derived from the discrete Fourier transform of the edge coordinates of the to-be-recognized moving object, where the edges are obtained through motion-based segmentation of the image scene. An adaptive threshold is used during the decision making process to indicate a match or identify a mismatch. Computer simulation on real world data proves the effectiveness of the proposed model. The proposed scheme is readily amenable to opto-electronic implementation.

The invariant image preprocessing of moment invariants generates an invariant representation of object features which are insensitive to position, orientation, size, illusion, and contrast change. In this study ARTMAP is used for 3-D object recognition of manufacturing parts through these invariant characteristics. The analog of moment invariants created through the image preprocessing is interpreted by a binary code which is used to predict the manufacturing part through ARTMAP.

A shift and rotation invariant neural network using interpattern hetero association (IHA) model is illustrated. To preserve the shift and rotation invariant properties, a set of binarized-encoded circular harmonic expansion (CHE) function at the Fourier domain is used as the training set. The interconnection weight matrix is constructed using an IHA model. By using the shift and symmetric properties of the modulus Fourier spectral, the problem of centering the CHE functions can be avoided. Computer simulations and experimental demonstrations are provided in which we have shown that the shift and rotation invariant properties of the proposed IHA neural net are indeed preserved.

The polarization properties of a TVT-6000 LCTV have been investigated. Mueller matrices of multiple ray paths through the TVT-6000 were measured for a single (typical) pixel, and through several pixels, using an imaging polarimeter. The TVT-6000 was characterized as a function of applied voltage and angle of incidence. From the Mueller matrices, the spatially dependent retardance, diattenuation, and depolarization are calculated and displayed as topographic maps. In another set of measurements, the LCTV is illuminated with a plane wave, and the spatial distribution of polarization in the Far Field Diffraction Pattern is measured in Mueller matrix form.

Some hybrid (analogue—digital) optoelectronic image processing system are considerecL Optical correlation processors with the pro— grammable trans-Formation core forming are proposed. Optical processor operation in the matrix—vector multiplication mode is discussed. A comparision analysis of optical image processors and digital universal and matrix calculators is carried outS.

Mathematical description and algebraic interpretation of recursive algorithm in optical binary image processing are carried out. The recursive optoelectronic processor configuration is described and experimental results are presented.

As applied to build up the real-time holographic correlative processing system it is shown the opportunity of the creation of the photo-thermoplastic media-based optical (holographic) read-only memory with high storage capacity (to 10 000 of high-informative images with l000xl000 points) and also random-access memory. Optical memory is integrated with real-time holographic correlator. It can be used for real-time processing of high-informative images that increasgs the accuracy and the reliability of the systems with its using.

We examined the principles or designing optical inrormation processing systems based on the dmamic spatial and temporal properties o two—pulsed photon echo signals. Also discussed are the pro— perties o1 the resonant medium taken as an operational filter o tern— poral and spational frequencies conditioned by the sensitivity or ex— citing light pulses to the Fourier—spectrum. It was shown that the ability o quantum resonant schemes to degenerate, maniresting itseir depending on the resonant response rorn the square amplitude o the second exciting light pulse, can be used ror simultaneous correlative processing or optical images formed at the same time moments. Dir— rerences between methods o1 simultaneous image processing obtained by means or the Fourier optics are also discussed in the paper.

The cause of the low recognition ability of the holographic systellis is the "hard" algorithm or activity for the sDecial a priori program. The necessity or such algorithm stipulated secific method of a coherent template-image formir]g.The problem can be solved by an intro'luction of the oreprocessirig operation into an ongoing incoherent and template channels in the real time scale, arid by means of synthesis of the template-image with the help of a computer. The .rtlcle deals with the scheme of a holographic processor, ex— tending possibilities of such approach, and the principles. of its design reasization are cpnsidered. The scheme, the photos of the working model and the technical data, received during the engineering tests are also given.

The method of the numerical evaluation of the influence of the geometrical distortions on the image correlation without the limitation on the type of the distortions is discussed. The main types of the distortions are described. The results of the theoretical modeling and experiments on the holography correlator are presented.