The optical correlator is applied to a robotic vision, tray-picking problem. Complex matched filters (MFs) are designed to provide sufficient optical memory for accepting any orientation of the desired part, and a multiple holographic lens (MHL) is used to increase the memory for continuous coverage. It is shown that with appropriate thresholding a small part can be selected using optical matched filters. A number of criteria are presented for optimizing the vision system. Two of the part-filled trays that Mendelsohn used are considered in this paper which is the analog (optical) expansion of his paper. Our view in this paper is that of the optical correlator as a cueing device for subsequent, finer vision techniques.

In this paper we continue our investigation of the application of matched filters to robotic vision problems. Specifically, we are concerned with the tray-picking problem. Our principal interest in this paper is the examination of summation affects which arise from attempting to reduce the matched filter memory size by averaging of matched filters. While the implementation of matched filtering theory to applications in pattern recognition or machine vision is ideally through the use of optics and optical correlators, in this paper the results were obtained through a digital simulation of the optical process.

A microcomputer-based real-time programmable optical pattern recognition system utilizing two magneto-optic spatial light modulators (MOSLM) with a liquid crystal light valve (LCLV) is presented. The proposed system has the advantage of high speed processing capability. Since the proposed system utilizes a microcomputer-based concept, it offers the advantages of flexibility and versatility. With reference to the joint correlation property, the system is capable of performing cross-correlation with various reference image patterns in real-time. Thus this hybrid optical architecture can be applied to image correlation with a wide variety of reference images. With appropriate design of interfacing units, hardware and software, the system is capable of learning new patterns, once it has encounted. Simple experimental demonstrations of this proposed system is provided.

For automation and robotics in the Space Station era, NASA's Johnson Space Center is pursuing several means of synthetic vision. The optical correlator is one such. The deformable mirror device (DMD) of Texas Instruments will form the basis of the first correlator in this project. In-house and contracted effort is being used. Initial in-house activities will concentrate on an impulse deconvolution technique and on a programmable retina for spatial remappings of an image prior to correlation. The retina will permit a form of edge extraction and other primitive operations. Additionally, it will be used as a research tool for assisting persons with low vision.

Two dimensional correlation for performing image recognition is one of the earliest examples of optical information processing suggested by Vander Lugt and it still remains one of the most promising application areas for this technology. The reasons optical image correlators have not found to this day widespread usage in pattern recognition systems fall into two categories: algorithmic limitations and lack of devices. Correlation-based pattern recognition algorithms have well known limitations such as scale and rotational sensitivity. Over the years algorithms based on optical correlation have been developed, which show promise for overcoming some of the algorithmic limitations, however they have not been put to a real practical test because of the lack of two dimensional spatial light modulators that are necessary for the implementation of the classic Vander Lugt correlator. Specifically, two separate two dimensional SLMs are needed, one for recording the input image and another for the reference.

We have recently presented (SPIE Proceedings, Vol. 579, pp. 215-224, 1985) a new hybrid optical/digital technique to compute the geometric moments of an image using its Fourier plane intensity. In this paper, we present the results of a simulation of this system. The parameters investigated include: input symmetries, sampling interval in the frequency plane, the order of the differentiating FIR filter, the detector area, frequency plane detector dynamic range and additive noise. An example is also presented to illustrate the applicability of this approach in a typical pattern recognition problem.

Morphological shape transformations on binary imagery can be performed by optically convolving the input image with the desired structuring element (e.g., a disk) and thresholding the output. Photographic film can be used for thresholding, but better results are obtained with electronic scanning and hard-limiting. Erosions, dilations, and median filtering can be performed directly. Openings and closings require an image store and feedback. Results of simple experiments are presented.

Texture is one of the important image characteristics and is used to identify objects or regions of interest. The problem of texture classification has been widely studied. Some texture classification approaches use Fourier power-spectrum features, while others are based on first and second-order statistics of gray level differences. Periodic textures that consist of mostly straight lines are of particular interest. In this paper, a new approach based on the Hough method of line detection is introduced. This classification is based on the relative orientation and location of the lines within the texture. Experimental results will also be presented.

A brief review of optical image subtraction techniques recently developed is presented. We have divided them into two categories according to whether they can or cannot be performed in real time.

A technique for performing an on-line transformation of an object from a compact, complete 3-D model in one fixed orientation is described. The model consists of connecting planar polygons in 3-D space and is useful for symbolic, associative, and other Artificial Intelligence (AI) processing. Two potential architectures are presented in which this system can be used in an "intelligent" recognition system.

In this paper a new shape classification scheme is proposed which uses the image processing formalism of associative memory mappings. This scalar transform technique is applied to two-dimensional (2D) images. The shape description is the centroidal profile which is the radius as a function of arc length parametrization of the boundary. Other one-dimensional (1D) representations are also discussed. The scheme is applicable to both full- and partial-view recognition problems. The restoration of degraded images, either due to occlusion or other forms of information loss, is optimal in the least squares sense.

The Hopfield neural network model has recently been proposed as a method for optically determining the nearest-neighbor of a binary bipolar test vector from a set of binary bipolar reference vectors. We illustrate several drawbacks of this approach and introduce a new technique called direct storage nearest-neighbor (DSNN) algorithm to accomplish the same task. We provide a comparison of the two approaches and demonstrate the superiority of the proposed DSNN algorithm.

A method of using synthetic-discriminant-functions to facilitate learning in a pattern recognition system is discussed. Learning is accomplished by continually adding images to the training set used for synthetic discriminant functions (SDF) construction. Object identification is performed by efficiently searching a library of SDF filters for the maximum optical correlation. Two library structures are discussed--binary tree and multilinked graph--along with maximum ascent, back-tracking, perturbation, and simulated annealing searching techniques. By incorporating the distortion invariant properties of SDFs within a library structure, a robust pattern recognition system can be produced.

An iterative method is used to design distortion invariant correlation filters. Target images can be detected, for example, independently of their position, rotation, or scale. Optical correlation filters designed using this technique retain full position invariance. The filter design begins by finding the distortion invariant modes for a particular image. This paper reviews the basic design process for a filter that is position, rotation, and intensity invariant. The emphasis is on determining the practical utility of the proposed methods and demonstrating the filters experimentally. Numerical simulations demonstrate that the proposed method can detect targets that are buried in noise. Computer generated holographic filters are fabricated using an electron beam. Experimental results are in excellent agreement with analytical and numerical computations.

An algorithm has been developed which reduces the effects of (deconvolves) instrument-induced spatial crosstalk in satellite image data by several orders of magnitude where highly precise radiometry is required. The algorithm is based upon radiance transfer ratios which are defined as the fractional bilateral exchange of energy beween pixels A and B.

A typical geographical map consists of a large number of lines and symbols. These symbols and lines represent various physical entities and their spatial relationships. Design of a computer-based system to automatically extract information from a map and answer queries is a very challenging task. This requires an intelligent query processor and a very sophisticated image analysis system. One of the tasks of the image analysis system is recognition of symbols and lines. In this paper we describe simple line tracking and symbol identification routines. The routines are capable of tracking various types of lines (continuous, broken, intersecting) and symbols having simple geometrical shapes.

A new technique for the displaying of continuous-tone color pictures on a bi-level, tri-primary display or a display with limited numbers of colors displayable has been investigated. The significance of this technique is that it can provide pleasing and realistic images on a standard, low-cost, RGB graphic display, which is in abundant use in the personal computer community. Specific hardware has been constructed to acquire images from a video camera and generate these images on several personal computer (PC) color graphic system monitors. The time to process the digitized 24-plane, RGB image into the dithered 3-plane, RGB image is less than one second. Images have been created for several potential applications including a semiconductor failure analysis expert system.

The Lister Hill National Center for Biomedical Communications, the National Library of Medicine's research division, is currently engaged in studying the application of Electronic Document Storage and Retrieval (EDSR) systems to a library environment. To accomplish this, an EDSR prototype has been built and is currently in use as a laboratory test-bed. The system consists of CCD scanners for document digitization, high resolution CRT document displays, hardcopy output devices, and optical and magnetic disk storage devices, all under the control of a PDP-11/44 computer. Prior to storage and transmission, the captured document images undergo processing operations that enhance their quality, eliminate degradations and remove redundancy. It is postulated that a "pre-processing" stage that removes extraneous material from the raw image data could improve the performance of the processing operations. The processing operation selected to prove this hypothesis is image compression, an important feature to economically extend on-line image storage capacity and increase image transfer speed in the EDSR system. The particular technique selected for implementation is one-dimensional runlength coding (CCITT recommendation T.4), because it is an established standard and appropriate as a base line system. The preprocessing operations on the raw image data are border removal and page centering. After centering the images, which are approximately 6 by 9 inches in the examples picked, in an 8.5 by 11 inch field, the "noisy" border areas are then made white. These operations are done electronically in a digital memory under operator control. For a selected set of pages, mostly comprising title pages and tables of contents, the result is an average improvement in compression ratios by a factor of over 3.

The diurnally varying IR imagery obtained from different field of view (FOV) is studied for the purpose of Wiener's whitening technique to be implemented in the image domain for real time clutter rejection. The measured power spectral density (PSD) on a focal plane arrays (FPA) is derived in terms of the propagation of the correlation of the ground irradiance l_λ(ε): PSD(k) = < Iλ(ε₀)Iλ (ε₀-Hk) >, where λ is the IR wavelength, ε_o is an arbitrary point within the downward FOV, the angular brackets denote an ensemble average, and H is the height above the ground. A combined measure of both dynamic ranges and resolution line pairs that can explain the thermal contrast and the diurnal variations is derived from Rayleigh's visibility V(r) by means of Taylor series expansion and logarithmic derivative of the measured temperature fluctuation correlation function C(r)= < ▵ T (0) ▵ T (r)>, such that V(r) = (r/2)•(d log C(r)/dr). The measured PSD(k) at FPA reveals systematically a diurnal trend of the power law variation of all FOV sceneries, |k|^-D-1 where D = 0, 1, 2, 3. Consequently Wigner distribution W( k,x₀) is introduced to generalize the Wiener PSD(k) to include the nonstationary FOV ye. Spatiotemporal filtering for clutter-rejection is a Wiener and Wigner whitening procedure in the sense of whitening in matching with the FOV. Analytical scaling laws of clutter leakage ≈εⁿ/1^D+1 useful for FPA designs have been derived in terms of a step size FPA resolution parameter ε=k_cx_s/2, the filter differencing order n, and a projected ground correlation length L at the FPA l=(D₀H|λR²L, where k_c is the optical cutoff frequency (D₀/λf₀), expressed in terms of the lens aperture D₀, the focal length f₀, and the pixel separation distance x_s, and the slant range R. Finally, the optical bench flight scene simulator is implemented using stationary scene input at a fixed FOV, and the optical flow on the moving platform FPA is derived. An apparent scene velocity along the radial direction / from a fixed point (x₀,y₀) on EPA is generated from a stationary input scene, on the FPA v(x,y)=(f₀H/R²)Ulsin[(x-x₀)²+ (y-y₀)²)^1/2]where U is the platform velocity component in parallel to the ground.

A paradigm to reduce computational costs in analyzing time-varying images is proposed in this paper. Our model is a hybrid of three recent advances in computer science, namely, hierarchical data structures, parallel processing, and heuristic planning. A pipelined pyramid image structure is constructed in the model by continually converging incoming images into successively lower resolutions. The model also contains a set of processors which work concurrently and asynchronously on subimages at different levels of this pyramid structure. These processors initially watch for interesting features in the lowest resolution rendition, of the scene. Processors working on promising areas individually but coopeiatively proceed to progressively higher resolution levels according to a planning scheme. This distributed planning mechanism is afforded through a blackboard control structure which also permits a unified scene interpretation. The model has been implemented in a simulated distributed system.

A prototype fingerprint identification system for personal verification has been developed. This system has a unique identification strategy, in the "top-down" direction, to map the features registered onto a raw input fingerprint image. Feature extraction by various image processing methods is not required and features used to make the top-down directional identification are selected that are easily detected on a raw input image. Therefore, this strategy makes identification possible with a personal computer.

Beginning in 1977, Fairchild Weston Systems developed an electro-optical reconnaissance camera for the Mission Avionics Division, Avionics laboratory at WPAFB, Ohio. The goal of the Long Range Electro-Optical Reconnaissance System program was to penetrate heavy haze conditions while preserving requirements of coverage and resolution needed to perform reconnaissance tasks. Design constraints included worldwide operational capability, high resolution, high sensitivity and large area coverage. It was proven feasible to achieve these goals through the use of hybrid analog and digital processing. Haze removal, local area background subtraction, data compression and image reconstruction were demonstrated. This paper discusses mission requirements, the design, and presents some results of this thirty-six month advanced development effort.

This paper presents a unified analysis of a class of unitary transforms including the discrete Fourier, the Walsh Hadamard, the discrete Hartley, and the discrete cosine transforms. These transforms possess a common recursive property that allows us to obtain the next higher-order transform from two identical, preceding lower-order transforms. This recursive property eventually leads us to formulate a fast, efficient recursive algorithm for the discrete Hartley transform, from which the fast processing algorithm for the discrete cosine transform can also be obtained. Hybrid implementations using state-of-the-art integrated optics in digital format have also been proposed for such fast, efficient processing algorithms.

A novel method for two-dimensional pattern recognition and feature extraction, applicable to microprocessor-based vision systems, is presented which employs fractal geometric analysis. Fractal contour transformation and transform correlation techniques are discussed in relation to their effectiveness in classifying rotationally deformed images over a wide resolution range. vractal geometric analysis exhibits several attributes: 1. position-, size-, and rotation-invariance is preserved in the absence of image coordinate transformation, 2. invariance to out-of-plane rotation is exhibited over the range ±60° of broadside, and 3. out-of-plane rotation can be computed from imagery and quantified in terms of the fractal dimension. This work is supported by experimental verification of a ship silhouette recognition algorithm. Results are presented in terms of recognition ratio and computational load.

Synthetic discriminant functions (SDF's) for matched filters have potential use for pattern recognition. However, these filters have been plagued with a low signal-to-noise ratio (SNR); i.e., these filters have no trouble correlating very well with true targets, but very often give high (even major) correlations with false targets. In fact, numerical experiments by the author and others show that the standard recipe for manufacturing SDF's gives filters with an SNR close to 1.00, even on a training set of imagery which has been edge-enhanced and energy-normalized. The author has introduced a new recipe for manufacturing SDF filters. These filters have an SNR of over 7.37 against their training sets and have proved to be very accurate in picking targets from extremely cluttered backgrounds, in fact much more accurate than the filter made from the standard recipe. The purpose of this note is to concisely recall the author's new recipe for these filters and to report on recent numerical experiments in which the difference in performance between filters made from the new and old recipes is, if anything, even more pronounced.