A simple target which contains a wide variety of spatial frequencies has been developed for use in testing optical correlators. This target is designed to be complicated enough to provide an adequate test of many correlator and filter architectures, but yet simple enough to be used by researchers without the need for special equipment. Computer simulations and experimental results are presented for using the target in testing some correlator filter designs.

The TOPS optical correlator program will demonstrate the capability for optical processors to perform target recognition and tracking tasks. This paper describes the demonstration to be performed. Operational considerations such as system complexity, reprogrammability, and throughput are discussed. Growth of these type systems and long term prospects are examined.

The TOPS pattern recognition system operational requirements are determined and system trades are identified. Component performances and issues are identified for the input spatial light modulator, detector arrays, solid lens, and electronics. A design description for the TOPS system is discussed. The overall system computational capability is evaluated.

Under the guidance of the USAF Electronic Systems Division of Rome Laboratory and Dr. Joe Horner, Martin Marietta is conducting a 32 month Transfer of Optical Processing to Systems (TOPS) Optical Pattern Recognition program culminating in an automatic target recognition flight demonstration in early 1994. During the three phase TOPS program, four optical processing systems involving at least two spatial light modulator (SLM) technologies will be developed and tested. Martin Marietta and the U.S. Army Missile Command (MICOM) will conduct an extensive test program including environmental testing for a helicopter flight, a tower test program, and a helicopter captive carry flight test program at the Redstone Arsenal to recognize M60A2 tanks.

Two sets of turntable and test range imagery of an M60A2 tank were collected to include various elevation, azimuth, and range variations. The turntable imagery was used as the reference for generating binary phase-only filters which were evaluated with the test range imagery. Several sensitivity and filter design parameters were determined for single reference and multi-reference 'smart' filters. The distortion sensitivity parameters derived from the initial data collection were used to determine the final turntable and test range data collection requirements. The final data collection will be used to generate and evaluate a complete smart filter set designed to recognize and locate an M60A2 tank.

The TOPS Electronic Warfare channelizer program is an effort being conducted under the DARPA sponsored Transition of Optical Processing into Systems (TOPS) initiative. The objective is a high probability-of-intercept radar receiver for signals in dense environments. Channelization is the preferred receiver approach in interfering signal environments because it instantly sorts time coincident pulse signals. Channelization is performed optically, using acousto-optical techniques, since optics offers size, weight, power and cost advantages. This program is a three year channelizer development effort. Once developed, the receiver will be tested integrated with an existing electronic warfare system. The paper will review the program background. The electronic, mechanical and optical requirements will be addressed. The critical technical issues impacting this program will be analyzed.

The military has invested significant resources for automatically locating and identifying enemy threats and high value targets. Automatic Target Recognition (ATR) systems using the output of imaging sensors, such as video, and Forward Looking InfraRed (FLIR) are continually being developed. The achievement of adequate processing rates required to analyze the tremendous amounts of data from these sensors remains a critical task in bringing these systems to the field. The two Optical Processing Identification Demonstration (OPID I and OPID II) program efforts, in their use of an optical processor, address the needs for both advances in system algorithms and data throughput enhancement.

This paper provides an overview of the design and operation of Lockheed's portable optical correlator. The design issues for a compact correlator are given along with the performance characteristics of the Fourier transform lens. The liquid crystal spatial light modulators are described and the system level description of the optical correlator given.

The performance of an optical correlator is analyzed as a vision system for a manufacturing robot. The correlator architecture is mathematically modeled and a simulation of the system is undertaken using representative airplane parts. Of particular interest in this study is the effect of three-dimensional part data on a two-dimensional correlation process. The model accounts for the apparent changes in part dimensions as a function of z-axis components, position in a parts tray, and the angular field of view of the vision system camera. The impact of these changes on the level of the correlation function and background discrimination is then assumed. Complementing the analytical model simulation, an actual optical correlator system is arranged to operate with parts placed at various locations in the physical tray area in order to obtain their corresponding levels. It was found that, by the introduction of multiple filters, the representative 3-D parts can be detected and tracked within the area of the tray, even when apparent part dimensions changed over the field of view of the imaging camera. The results of this research indicate that the correlator vision system concept shows promise for future 3-D manufacturing robotics applications.

A liquid crystal television (LCTV) SLM's phase, amplitude, and polarization all influence our selection of operating curves for input and filter. With its continuum of drive voltage, the LCTV permits grey-level control in both locations. Selection of an optimum curve depends on expected variations in signal amplitude, presence of input scene structured noise, and other factors. Using modulators obtained from a commercially available projection LCTV, and with no specifically added input noise present, we have obtained laboratory results in which the ratio of peak intensity to correlator system noise exceeded 100:1. Unfortunately, it was massively inefficient to implement the algorithm which had been developed as the abstract for this paper was submitted. Fortunately, we have quite recently developed two insights that will speed it by several orders of magnitude, and we shall report that result in the future. We will also extend the work to include clutter objects and additive input scene noise.

A multisensor image analysis system that locates and recognizes realistic models of military objects placed on a terrain board has been demonstrated. Images are acquired using two overhead video sensors--a wide field, low resolution camera for cueing and a narrow field, high resolution camera for object segmentation and recognition. The red, green, and blue sensor information is fused and used in the digital image analysis. Small regions of interest located within the wide field-of-view scene by a high-speed digital cuer are automatically acquired and imaged by the high resolution camera. A high-speed statistical segmenter produces a binary image of any military object found within a given region and sends it to a computer-controlled binary phase-only optical correlator for recognition. Rotation, scale and aspect invariant recognition is accomplished using a binary tree search of composite binary phase-only filters. The system can reliably recognize any one of ten different objects placed at any location and orientation on the terrain board within ten seconds.

Liquid crystal televisions have become increasingly popular as low cost spatial light modulators. While the early devices suffered from poor resolution and low contrast, recent models compare favorably to the more traditional (and expensive) modulators. Most of the LCTVs are addressed through an RF interface to the external antenna or through a direct video input. This is a convenient interface for an input plane spatial light modulator where the scene information is provided by a video camera. Filters, however, are generally computer generated and the conversion to video can sometimes be an awkward step. A new drive circuit for the Epson Crystal Image video projector is currently under development at the U.S. Army Missile Command which allows pixel level control of the liquid crystal sandwich. This paper will discuss the proposed circuit designed to allow pixel by pixel control of the LCTV.

Quantex's electron trapping (ET) materials are investigated as an integrated spatial light modulator (SLM) at the filter plane in a joint transform correlator (JTC). The ET based spatial light modulator can detect visible incoherent light, store an image pattern, and modulate the incoming coherent infrared light. Such a JTC architecture eliminates the electronic-to-optical and optical-to-electronic conversion between the square law detector and the next SLM in a joint transform correlator.

We demonstrate a new optical correlator in which the correlation peak intensity is increased when the matched input object is moving. The basic configuration of the correlator is the same as a VanderLugt optical correlator consisting of a photorefractive crystal. The principle of this new correlator is based on the dynamic grating erasure property of photorefractive materials. The detail of this principle is described.

Much research effort has been directed to designing phase-only and binary phase-only filters that maximize the signal-to-noise ratio (i.e., the noise tolerance) or the peak-to-correlation energy (i.e., the correlation peak sharpness). In this paper, we propose a filter design algorithm for phase-only filters that produce correlation outputs that approximate classical matched filter outputs in a minimum mean squared error sense. Simulation results are presented to illustrate the use of the design algorithm.

The ternary phase-amplitude filter (TPAF) is by definition restricted to the modulation values -1, 0, and 1, thus comprising 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. The TPAF enables filter implementation with magneto-optic devices and these devices also can also be used for image input if gray scale scenes can be binarized while preserving good correlation performance. We provide simulation results addressing the comparative performance of TPAF correlation using gray scale and binary images derived from identical edge-enhanced version of original scenes. Experimental correlations verifying the results for the binary input cases are provided.

Binary phase-only filters (BPOFs) have generated much study because of their implementation on currently available spatial light modulator devices. On polarization-rotating devices such as the magneto-optic SLM, it is also possible to encode binary amplitude information into two SLM transmission states, in addition to the binary phase information. This is done by varying the rotation angle of the polarization analyzer following the SLM in the optical train. Through this parameter, a continuum of filters may be designed that span the space of binary phase and amplitude filters (BPAFs) between BPOFs and binary amplitude filters. In this study, we investigate the design of optimal BPAFs for the key correlation characteristics of peak sharpness (through the PCE metric), SNR, and discrimination between in-class and out-of- class images. We present simulation results illustrating improvements obtained over conventional BPOFs, and trade-offs between the different performance criteria in terms of the filter design parameter.

Extremely light efficient optical correlators can be built if both the input and filter are displayed on phase modulating devices. Problems may exist, however, since changes in the input scene intensity cause changes in the phase encoding. An analysis is presented here for the VanderLugt correlator architecture. Experimental results comparing the correlator response to phase-encoded and amplitude-encoded inputs are also presented for varying input scene intensities.

Even in the absence of input noise, there is no guarantee that correlation peaks resulting from some filters such as the binary phase-only filters (BPOFs) will be at the origin when the input is the reference object centered. Simulation results are included that show this peak shift is usually negligible.

Our research group at Carnegie Mellon University has devised a large number of new distortion-invariant optical correlation filters. We briefly review our hierarchical inference approach to scene analysis, our minimum noise and correlation energy (MINACE) filters, the use of morphological hit-miss (rank-order) and MINACE filters for detection, and the use of MINACE filters with different control parameters c for all three levels of scene analysis (detection, recognition, identification). We then detail new generalized rotation-invariant filters with initial test results.

Several versions of synthetic discriminant functions have been proposed over the last twelve years. This paper describes the results of tests on the 'Minimum Squared Error SDF,' (MSE- SDF), a recently proposed variation, with some specific unique performance detection capabilities.

This paper describes work being performed at Martin Marietta to synthesize and implement robust, distortion-tolerant composite filters to perform automatic target recognition (ATR). Two main topics will be discussed. The first part of the paper outlines two filter design methods for trading between distortion tolerance and false alarm rate. In the second part, we will explain the real-time digital implementation of a multiclass ATR system using composite filters in Martin Marietta's massively parallel computer, the Geometric Arithmetic Parallel Processor (GAPP^TM).

A very common operation in optical processing is cross correlation of an input scene with a filter to extract information from the input scene. For instance, in the identification mode a correlation surface should contain a high value at the location corresponding to the position of the object in the input space and significantly lower values at all other locations. The performance of this operation is unfortunately degraded when the input contains noise, or the object is distorted, rotated, shifted, occluded, or changes in aspect. A corrupted input requires that the filter have some generalization capability. This capability is usually gained at the expense of the peak sharpness and height. Thus, filter architectures that mold the correlation surface to effectively trade peak sharpness and generalization need to the considered. This paper considers several filters of the SDF family with regard to their ability to shape the output correlation surface.

An optimal quadratic filter is developed to detect a target signature distorted by a linear transformation group (e.g. translations, scaling, etc.). The filter is constructed from the invariant subspaces of the transformation group. This guarantees an equal response to all target images. A maximum signal-to-clutter ratio is achieved with a weighted combination of invariant subspaces. Examples are discussed for some one-parameter subgroups of the projective transformation.

We describe how the synthetic discriminant function (SDF) can be employed in sensor/data fusion. This is accomplished by exploiting the isomorphism between the SDF formalism and the parallel distributed processing (PDP) architecture of neural networks.

In the past, several different approaches to Synthetic Discriminant Function (SDF) filter design have been proposed. These include: conventional SDFs which control the correlation values at the origin, Minimum Variance SDFs (MVSDFs) which minimize the noise sensitivity of the filters, Minimum Average Correlation Energy (MACE) filters which maximize the peak sharpness, and Linear Phase Coefficient Composite (LPCC) filters which use phasor addition and subtraction for inherent class discrimination. In this paper, we introduce a new family of SDF filters of which all the above are special cases. Each filter in this family is characterized by two parameters (alpha) ₁ and (alpha) ₂. Various choices of ((alpha) ₁,(alpha) ₂) lead to above special filters. For example, (alpha) ₁ equals 1 and (alpha) ₂ equals 0 leads to MACE LPCC filters which are hybrid versions of MACE and LPCC filters. This family of filters is evaluated using the Minimum Probability of Error (MPE) criterion and a data base of aircraft images. These simulation experiments confirm the superior performance of this filter family. Also, we observe the interesting result that the MPE is at its lowest not for one of the four special filters listed above, but for a combination of them.

We propose a new method to find the proper center of the circular harmonic filter using simulated annealing. The new method is two orders of magnitude faster than the previous method. The proper center is a chosen expansion center which ensures the maximum peak of the correlation output to be at its center, thus a truly shift and rotation invariant pattern recognition system. The proper center is to be used in a new synthesis of filters: the geometrical approach of multiple circular harmonic filters. In this approach, the distances between the correlation peaks in multiple correlation outputs are used as features for pattern recognition. This method has good discrimination capability and can be implemented for practical applications.

The design of a highly selective synthetic discriminant function (SDF) filter is described for implementation in an up-dateable correlator based on a non-degenerate four wave mixing interaction in the photorefractive crystal Bismuth Silicon Oxide. The coefficients of the SDF cross-correlation matrix are determined for each member of the training set by modeling the high-pass filtering effect generated by a Fourier transform hologram written to the BSO with a relatively weak reference beam. Simulation show that the SDF so produced results in a correlation peak localization and discrimination ability comparable to that of the MACE filter. The filter is realizable as an all positive real-valued space domain image capable of direct input to the correlator via an SLM. Simulation results are presented for the orientation- independent, shift-invariant recognition of a test industrial component. The filter's discrimination ability against another test component, not included in the filter training set, is also examined.

In an optical correlator, binary phase-only filters (BPOFs) that recognize objects that vary in a nonrepeatable way are essential for recognizing objects from actual sensors. An approach is required that is as descriptive as a BPOF yet robust to object and background variations of an unknown or nonrepeatable type. We developed a BPOF that was more robust than a synthetic discriminant function (SDF) filter. This was done by creating a filter that retained the invariant features of a training set. By simulation, our feature-based filter offered a range of performance by setting a parameter to different values. As the value of the parameter was changed, correlation peaks within the training set became more consistent and broader. In addition, the feature-based filter was potentially useful for recognizing objects outside the training set. Furthermore, the feature-based filter was more easily calculated and trained than an SDF filter.

An optical neural network based upon the Neocognitron paradigm is introduced. A novel aspect of the architectural design is shift-invariant multichannel Fourier optical correlation within each processing layer. Multilayer processing is achieved by iteratively feeding back the output of the feature correlator to the input spatial light modulator and updating the Fourier filters. By training the neural net with characteristic features extracted from the target images, successful pattern recognition with intra-class fault tolerance and inter-class discrimination is achieved. A detailed system description is provided. Experimental demonstrations of a two- layer neural network for space objects discrimination is also presented.

A hybrid opto-electronic neural network using light emitting diodes, fixed pattern masks, and linear photodetectors is proposed and analyzed. The compact design, low power requirement, and high computational rate make this approach attractive for a variety of optical pattern recognition problems.

Exemplar-based neural net classifiers enjoy extremely rapid learning procedures and are particularly suitable for analog optical hardware implementations. The winner-take-all (WTA) network is a key component in exemplar-based neural net classifiers as well as in optical competitive learning architectures. In this paper, we present an optical WTA network based on novel electron trapping (ET) materials. The mathematical model has been modified for the optical implementation. All the neuron operations required by the WTA network such as self- excitation, lateral inhibition and thresholding, are performed by a single ET device.

We present a novel neuro-optic system for ATR based on an advanced holographic technique. This system has a spatially recorded high density holographic matrix which works in conjunction with a 2-D spatial light modulator to perform highly parallel inter-pixel operations. Because of the high degree of parallel processing necessary for real-time performance and the need for high interconnectivity and large memory capacity, an optical neural network approach to ATR fulfills more than the basic requirements for an ATR system and has the potential to outperform, or at a minimum supplement, an electronic neural network approach. Our neural network is based on inter-pattern association model and interconnection network is implemented using N⁴ hologram recording. In this paper, computer simulation as well as optical implementation of the neural network is presented. Computer simulations is used to implement neural network with 64 X 64 neurons for rotation invariance and weak scale invariance in a controlled environment. For the optical implementation we used 32 X 32 fully interconnected neural network to test noise robustness.

This paper describes the work on optical neural network-based IR target recognition. The experimental system consists of the fabricated optical synaptic device with two stacked layers, i.e., GaAs-LED and Si-PD, with the interconnection matrix patterned on the PD surface. The experiment for four patterns representing small moving targets shows basically satisfactory results.

We have stored up to 5,000 holograms of high-resolution images (320 X 220 pixels) within a single crystal of Fe:LiNbO₃. This storage capacity permits the construction of a processor which can compute all the inner-products between any input image and the stored images simultaneously. Without having to retrieve information from a secondary mass memory, overall computation speed of inner-products can be very high. In this paper, we will describe our architecture; address performance issues; and present experimental results on reconstruction of holograms and inner-product computation.