Optical pattern recognition using diffraction pattern sampling has matured to the point that sophisticated laboratory, factory and scanner systems have been developed. In this paper we review the basic optical theory and a brief history of systems that have been developed. Two state-of-the-art systems will be discussed in greater detail. One is a laboratory configuration, while the other is a high-speed scanner. Recent results obtained utilizing the laboratory system to automatically classify terrain types will be presented. These results illustrate some of the limitations pure diffraction pattern systems have when used for texture measurement. Finally, the implication these results have on future directions of diffraction based systems will be discussed.

The results are provided of the second phase of a program to determine the feasibility of fabrication of a coherent optical pattern recognition system to locate all pages in a microfilm data base on which a given key word occurs. Results obtained using an optical frequency plane correlator system with weighted matched spatial filter synthesis in a scaling correlator topology have been most promising. By proper matched spatial filter synthesis and system design, correlation degradations due to various expected differences between the input and reference functions in this optical word recognition system have been decreased to acceptable levels.

This paper introduces a spatial pattern recognition processing concept involving the use of spectral feature classification technology and coherent optical correlation. The concept defines a hybrid image (data) processing system incorporating both digital and optical technology. In this concept, the hybrid instrument provides simplified pseudopattern images as functions of pixel classification from information embedded within a real-scene image. These pseudoimages become simplified inputs to an optical correlator for use in a subsequent pattern identification decision useful in executing landmark pointing, tracking, or navigating functions. In this concept, real-time classification is proposed as a research tool for exploring ways to enhance input signal-to-noise ratio as an aid in improving optical correlation. Although such a system has not been developed or tested, the approach could be explored with developing technology, including a current NASA Langley Research Center technology plan that involves a series of related Shuttle-borne experiments. A first-planned development phase includes a feature classification experiment. The experiment, Feature Identification and Location Experiment (FILE), is undergoing final ground testing, and is scheduled for flight on the NASA Shuttle (STS2/flight OSTA-1) in 1980. FILE will evaluate a technique for autonomously (in-real-time) classifying Earth features into the four categories of bare land; water; vegetation; and clouds, snow, or ice. A second experiment is being designed to test a technique for autonomously discriminating between clouds and snow. Beyond feature identification/classification and cloud detection/discrimination, the onnoing technology development leads to capabilities for pointing instruments to predetermined sites, reacquiring Earth features or landmarks, and tracking features such as coastlines or rivers. Concepts are discussed relative to emerging technology,and directions for future research are indicated.

Optical power spectral (OPS) analysis has demonstrated significant advantages for rapidly classifying simple patterns from aerial transparencies. However, as the pattern sets become larger and more complicated, the required software is more complex, the equipment requirements more demanding, and the processing time increases. Additionally, potential ambiguity of the OPS data due to loss of phase information places some unknown upper limit on the information content in OPS. Integration of other pattern recognition data which supplements the OPS analysis leads to an increased pattern classification potential. The optical/ Digital system discussed integrates a photo diode array detector in the image domain with the OPS system. Hierarchical strategies employ initial OPS sampling and preliminary classification of patterns of potential interest. Space domain sampling and classification is only initiated when cued from the OPS samples. This system leads to a wide range of experimental capabilities and operational potentials. The paper discusses system configuration, experimental results, and future plans.

The Inspectron is an instrument which will perform an automatic optical inspection of the etched circuitry on the individual layers of a multilayer printed circuit board. The concept of this instrument is unique in that it does not compare the PCB under test with a master or with computer-stored data. Instead, as the board is optically scanned, a small area is re-imaged onto a detector array. The detector signals, after digitization, are fed into high-speed logic circuitry which is programmed to distinguish between the appearance of a good board and an error. An 8 x 10 inch board can be inspected for line width, line spacing, line breaks, excess copper, and voids in about a minute. The Inspectron can also inspect pads for completeness and ground planes for shorts.

The holographic method can be used to determine the statistical characteristics of a slightly diffusing surface, the local rugosity ( mean quadratic variation ) of a surface and the lateral mean dimensions of diffuzing facets ( dimensions of the correlation zones). This is done by modifying Vander Lugt's method, so that the holographical restitution will recover its linear characteristic ( the linearity is lost when the darkness variations on the photographic plate are too high ).

A classification of multi-sensor imagery from the sensor's point of view is advanced. From this treatment, the statistical and deterministic contributions to a multi-sensor image correlation process are more clearly seen. The optimum preprocessing operation for several cases of multi-sensor image pattern recognition are noted and the use of weighted matched spatial filter synthesis as a one step optical pattern recognition correlator is described. Theoretical formulation and experimental verification of the result that edge enhancement preprocessing is not always optimum in a multi-sensor optical image pattern recognition system are presented.

In using matched filtering techniques to recognize and count species of diatoms, the two basic problems are the sensitivity of the system to size and to orientational variations between individuals of a given species. Since diatoms seem to be randomly oriented on the slides one reasonable approach to overcoming these difficulties is to make the filter matched to some average nember of the species and then to use statistical information to estimate the actual size of the population. This technique is applied to two populations of diatoms. Experimental results are presented and discussed.

The role of coherent optical pattern recognition techniques in image registration for track assembly and target discrimination applications is addressed. Use of optical pattern recognition on visible and simulated infrared mosaic imagery is found to have adequate accuracy to merit further attention. A frequency plane correlator using weighted matched spatial filter synthesis techniques was found to be a most optimum and flexible optical image registration system.

A hyperspace description of the multiclass (multiobject) pattern recognition problem is advanced. Average filters composed of linear sums of orthonormal basis functions are found to provide the necessary discriminant hypersurfaces. Optical correlation using weighted matched spatial filter synthesis is used to determine the basis functions and their linear weights. The average filters are then assembled on a digital computer and used in an optical frequency plane correlator. Initial problem formulation theory and simulation results are included for an infrared tank pattern recognition problem.

We used the optical Fourrier transform to point out faults and geological irregularities in the north of France. Using mask, we eliminate linear characteristics in chosen directions and bring out those having complementary directions. We make quantative measurements of the spatial distribution of the fractures and of the direction of the isointensity curves. We establish histograms of the spatial distribution of the fractures and curves. We develop computer-generated synthetic filter holograms. We use these filters to perform mathematical operations: derivative more complex operations. We are developing an interface in the Fourrier plane from optical to numerical treatment for specific geographical application.

Incoherent light can be subdivided into spatially and temporally incoherent light. Both types of light allow one to perform correlation operations for character or pattern recognition. We will briefly review methods in spatially incoherent light, but concentrate on wavelength-multiplexing with temporally incoherent light, i.e. light emitted from a white light source.

For the past several years, we have been conducting research on hybrid systems that combine nonconventional incoherent optical spatial filtering systems with digital or analog electronics to allow bipolar (and, if desired, complex valued) spatial filtering of self-luminous or incoherently illuminated object distributions [1-4; see also 5-9]. As has been emphasized in previous papers, such systems are free from many of the input problems that characterize coherent spatial filtering systems and are free from the blemish noise characteristic of coherent optical systems in general. Hybrid systems with a bipolar capability are necessary if incoherent spatial filtering is to be applied successfully to a wide range of image processing problems including feature enhancement and pattern recognition. We have recently implemented an experimental system suitable for a wide range of experiments in this area. This paper reports on preliminary tests of this system. We begin with a brief analytical review of incoherent spatial filtering, then consider specific hybrid implementations for bipolar spatial filtering.

In contrast to conventional sampling, implicit sampling permits the unique representation of a bounded bandlimited bipolar function in terms of a set of positive real numbers. This is important for noncoherent optical processing as a means of handling bipolar input functions without suffering the non-coherent light buildup and associated loss of contrast in the output plane of the processor usually encountered in other approaches. We will describe recent results of applying implicit sampling in optical correlation and Fourier transformation that illustrate the utility of implicit sampling in optical computing.

A novel approach to coherent optical pattern recognition is described. It does not utilize matched spatial filters and optical correlation methods. Rather, a 2-dimensional input function is described in terms of its absolute normalized invariant moments. Comparison of these moments enables one to determine the presence of a given object independent of geometrical distortions. A parallel optical processor with a special mask is used to generate all moments for a 2-D scene in parallel. A simple digital post processor calculates the actual absolute normalized invariant moments with the high dynamic range necessary. Our initial work reported upon here has concentrated on: the optical generation of individual moments, how the individual moments vary, the dynamic range requirements of the system and how they may best be met in a hybrid optical digital topology,and methods to generate all moments optically in parallel.

Matched filtering is the key concept in coherent optical pattern recognition. Unfortunately, matched filters are more sensitive to some things (nonconsequential variations in size, rotation angle, etc.) and less sensitive to other things (differences between object classes) than we might desire. The generalized matched filter is a filter tailored to the specific requirements of the task. It contains the matched filter as a special case.

A coherent optical method of performing pattern recognition has been studied which is invariant to selectable pattern features. Two examples of pattern features frequently encountered in pattern recognition tasks are scale and rotation. A recognition scheme which is invariant to these features is able to detect an object independent of its size and angular orientation. In our case, the recognition scheme not only is invariant to these specific features, but can also be used to measure them.

This paper reviews the operation of the hybrid field-effect liquid crystal light valve (LCLV) and summarizes the performance useful for real-time coherent optical data processing. The light valve is basically a high resolution optical-to-optical image converter. The device embodies a CdS photoconductor, a CdTe light-absorbing layer, a dielectric mirror, and a biphenyl liquid crystal layer sandwiched between indium-tin-oxide transparent electrodes deposited on optical quality glass flats. The input image is directed onto the photoconductor to reduce the impedance of the photoconductor, thereby switching the ac voltage that is impressed across the electrodes onto the liquid crystal to activate the device. The ac operation ensures long operating life for the device. The liquid crystal is operated in a hybrid field-effect mode. It utilizes the twisted nematic effect to create a dark off-state (voltage off the liquid crystal) and optical birefringence to create the bright on-state. The liquid crystal modulates the phase of the coherent readout light. By an additional analyzer, an intensity modulation is created. The current performance of the light valve is reported with the present performance for resolution, interharmonic distortion, signal-to-noise ratio, contrast ratio, and response time.

Optical Pattern Recognition results using the thermoplastic device as a complex filter in both the Vander Lugt scheme and the joint Fourier transform scheme are reported. Both the glass and tape substrates have been used for the correlation experiments with success. An extension of the joint Fourier transform technique to ambiguity processing is also discussed.

To determine the proper input format to utilize the real-time two-dimensional coherent imaging devices for antenna data processing, the properties of a raster scanned wideband signal are studied. The analysis of the ambiguity function leads to the possibility of two-dimensional processing of raster signals. Holographic filters are designed for multidoppler matched filtering and the array-pattern synthesis of 2-D wideband antenna-arrays, with multiplexing of rastered channels. Complete processing using a single holographic filter is presented experimentally for the case of a Sonar-doppler array antenna, with correlation detection and extraction of the three target parameters range, doppler, direction.

A computer model, with laboratory confirmation, was constructed to serve as a tool for analysis of RF signal generation and compression using acousto-optics. The emphasis of the model was on incorporation of realistic hardware and noise parameters. Application discussed here is intended to point out the general nature of the model and to review an example application to linear frequency modulated radar waveforms.

Pattern recognition systems that use optical systems are primarily oriented at using the inherent parallel processing capability of optical systems. Recent advances in marrying the parallel processing of optical processors and serial digital processing are optimizing the best of both approaches and providing a methodology for learning or adaptive processing systems. Examples of these hybrid system developments will be shown along with a generalized approach to modeling an adaptive processor with Weiner-Kolmogorov prefiltering.