The operation of a deformable mirror device (DMD) as a Fourier plane phase modulating filter is described. An analysis of the optical characteristics of the DMD elements as phase modulators is summarized. Analytical and experimental results indicating the existence of a quasi-phase-only operational mode are presented. These results are used to specify the mirror deflection required to implement a binary phase-only image correlation operation. An optical correlator system is implemented using the DMD Fourier plane filter and experimental results from this system are compared with computer simulations of the correlator operation.

We propose a real-time optical edge enhancement technique based on the polarizarion modulation of liquid crystal televisions. Experimental results and discussions are presented.

We consider various types of optical computer generated holograms (CGHs) using various encoding techniques. The application chosen is their use as optical matched spatial filters (MSFs) in an optical frequency plane correlator. This differs significantly from the use of CGHs for image reconstruction. Optical laboratory data is presented comparing the major types of CGHs on the same input data, using binary CGHs. The various issues associated with this use of CGHs are noted and quantified with optical laboratory data.

An optical correlator guidance technology demonstration program has been designed to illustrate the usefulness of an optical correlator in autonomous target recognition, tracking and vehicle guidance. The program will start with optical correlator systems which can recognize and track a cooperative target with a single holographic matched spatial filter, and will conclude with a system capable of autonomous recognition and tracking of a realistic target.

Optical correlators are subject to point-spread and field aberrations in both the Fourier (filter) plane and the correlation (output) plane. These aberrations can be both quantitatively and qualitatively different within a given system. Analysis of a simple correlator using identical spatial light modulators in the input and filter planes shows that even a well-corrected system of this type will be limited in performance by diffraction effects.

Simulations of the standard optical correlation process show that difficulties arise if the inputs are realistically modeled to be an aperture in a plane. These difficulties are especially grave if the inputs are targets in a background with substantial amplitude. Simple examples show that these difficulties cannot be solved by edge enhancement techniques or by suppression of low frequency terms in the Fourier plane. The purpose of this paper is to give two new variants of the standard optical correlation process which solve this aperture problem. This assertion is empirically supported by a great deal of testing over numerous data sets. These two variants are to be used with phase-only or binary filters implemented in an addressable spatial light modulator in the Fourier plane. Each variant requires its own careful nonstandard design of the phase-only filters. The direct construction of phase-only filters developed previously by the author can and must be used to make filters suitable for these variants. Simulations suggest that the resulting filters can be packed with a great deal of information, are stable under perturbations in the training set, have a very low false alarm rate, and are extremely insensitive to backgrounds consisting of high variance, high mean, Gaussian or uniform noise. These two variants give superior results even when targets are in a zero or low amplitude background.

Composite filters constructed with synthetic discriminant function algebra and binary phase-only Fourier transforms of training images are demonstrated through digital simulation. These are shown to provide high multiplicity and S/N, with unusually low scatter in S/N parameters compared to filters constructed with input plane images.

Using Page Oriented Holographic Memories (POHMs) optically addressed Spatial Light Modulators (SLMs), joint transform correlators, 2D or 1D acousto--optic cells, and optically addressable RAMs we can produce a massively parallel optical data base management system.

The architecture and experimental results for an incoherent acousto-optic image correlator suitable for real-time applications are presented. In the basic architecture, each time a line of the raster-scanned input image is fed into the acousto-optic device (AOD), all rows of a digitally stored reference image are read into the system using an array of light emitting diodes (LEDs). Thus, the required two-dimensional correlation is performed as a series of multi-channel 1-D time-integrations in x (performed in the AOD) combined with a multi-channel correlation in y (perpendicular to the AOD axis) using a modified CCD. The LED array and detector modifications which markedly increase the dynamic range are discussed as well as the correlator design. Further, a novel memory for storing the reference object is described for rapidly changing templates. Experimental results indicate the architecture is useful for applications in the areas of character recognition and target identification.

In this paper, we compare the performance of the bipolar joint transform correlator to the continuous phase-only matched filter correlator, the binary phase-only matched filter correlator, and the classical correlator. The pattern recognition systems will be compared in the areas of correlation peak, autocorrelation peak to sidelobe ratio, autocorrelation bandwidth, and discrimination sensitivity.

Computations for 3-D vision are greatly simplified by looking at the world through polar exponential sensor arrays. Rotation and zoom invariance, and wide field of view with reduced pixel count, are well known benefits of such arrays. More subtle, equally powerful properties applicable to the interpretation of 3-D images, are described in this paper. Computations involving scaling, such as perspective and optic flow, are reduced to additive operations by the implicit logarithmic transformation of image coordinates. Expressions for such computations are derived in this paper and applied to sensor design examples. Advantages of polar exponential arrays over x-y rasters for binocular vision are also discussed.

Rigid objects can be modeled by quadric or planar polyhedra. This representation allows low-level data driven algorithms to determine the identity and pose of an object. Before an object can be matched with an internal model, however, its image must be first segmented into basic components, such as smooth patches. This paper describes a low-level parallel range data analysis system that uses input range data to characterize the underlying scene in terms of the visible surfaces, using simpl0e, invariant features to generate a view independent description of the scene. The central components of the system are a parallel segmenter and a parallel surface fitter. The segmentation algorithms are based on the estimation of surface curvatures. Surface fitting allows for pose estimation and recognition of objects in range scenes. Mathematically, surface fitting can be formulated as an overdetermined linear system which can be solved in the least-square sense. Because of numerical stability and ease of implementation, the QR-factorization using the Givens transformation is suited for the parallel solution of overdetermined systems. We discuss the implementation of the range analysis system on a distributed-memory hypercube parallel computer.

In model-based recognition of vehicles, wheels can play a key role. Certainly, they are the most prominent features of a vehicle. However, finding wheels in an image is a difficult task. Generally, a wheel appears as an ellipse in an image. An obvious way of finding ellipses is to use the Hough transform. The difficulty is that the search space is 5-dimensional. This curse of dimensionality is with us no matter what method we use to search for the ellipses. We use a stereo pair of images to reduce the search space. The idea is to determine from the stereo images the 3-D orientation of the plane containing the wheels, and then apply an appropriate transformation on either of the two stereo images such that in the new image the wheels will be circular. The search space is then only 3-dimensional. In this paper we describe this approach in detail and show some experimental results.

As the possibility of stereo controlled robots becomes a reality, the need to transmit the stereo pair of images to a ground station or space station for man-in-the-loop supervision will be a necessity. The complexity of transmitting stereo images by coding the preprocessed disparity is discussed in this paper. The approach demonstrates the quantization, modulation, and reconstruction of the stereo images. Results indicate the accuracy of reconstruction in terms of mean-square-error criterion as a function of the signal-to-noise ratio. Key research issues of interpolation from sparse disparity maps and reconstruction of the stereo pairs in the presence of spatial noise are presented. It is concluded that stereo reconstruction is possible and the noise constraints are given.

NASA's Johnson Space Center (JSC) has created the specifications for a new kind of image processing machine, a video-rate coordinate remapper. JSC has contracted with Texas Instruments for its detailed design and construction. Previous image processing machines for machine vision typically have done point operations without changing the geometry of the image. The JSC/TI Programmable Remapper offers complete flexibility in changing the coordinates of an input image at video rates. Presented here are the main capabilities and some of the applications of such a device.

Polar exponential sensor arrays have been shown to be advantageous over conventional x-y raster imaging sensors when wide field of view, high central resolution, and rotation and zoom invariance are desirable. Applications include spacecraft docking, tracking and station keeping, and mobile robot navigation. Optimal designs minimizing sensor configuration and computational requirements are critical in such mobile applications. This paper analyzes geometric parameters of polar exponential arrays, and their relation to 3-D sensing precision requirements which drive sensor design. Resolution, field of view, and perspective projection are discussed. A method for smoothly patching the "blind-spot" singularity in the sensor with a uniformly high resolution "fovea" is presented. A software testbed for simulation of the sensor geometry and mappings is described. The user can specify any resolution and field of view. Imagery generated by this simulator efficiently illustrates the geometric effects of various choices of parameters, thereby providing a design tool for construction of sensors.

Stereo information can be obtained using a single translating camera. The analysis of a sequence of images obtained using such a camera configuration may be facilitated by using a mapping similar to the retino-striate mapping in mammals, which is the Complex Logarithmic Mapping (CLM). The mathmatical properties of this mapping make it ideal for object recognition, depth determination, and dynamic scene segmentation. There is widespread interest in the military, space, and industrial communities in algorithms for analysing images from a moving camera. Applications include autonomous vehicle navigation, space station construction and maintenence, and robot arm control for manufacture and inspection. In all of these cases, at least rough ego-motion information is available. Using this information to chose the origin of the mapping allows us to use important characteristics of optical flow without the need to calculate it. This technique, called Ego-Motion Complex Logarithmic Mapping (ECLM), can be used to calculate depths. Since the mapping is performed on a digitized image, a direct implementation of the CLM results in an image that contains less information than the original image. Computer vision methods for static scene analysis rely mainly on point, line, and region techniques. We explore how some of these algorithms perform in ECLM space and how they can be used for depth determination.

Much of vertebrate midbrain and mammalian cortex is dedicated to two-dimensional "maps" in which two or more stimulus parameters are encoded by the position of neural activation in the map. Moreover, there are a large number of such maps which interact in an unknown fashion to yield a unified perception of the world. Our research program is based on studying the structure and function of brain maps. In the present paper, we review a recently constructed system of computer aided neuro-anatomy which allows high resolution texture mapped models of cortical surfaces in two and three dimensions to be displayed and manipulated. At the same time, this work demonstrates some of the basic geometric patterns of architecture of the primate brain, such as columnar and topographic mapping.

A means for obtaining both a rectangular and a log-spiral representation of images is examined. A method which utilizes a single CCD imager is examined in detail. The method uses a pixel combination process to form the log-spiral representation. The additional circuitry which is necessary to perform this combination process is discussed. The combination process used forms a log-spiral representation which is distorted. A means for quantifying this distortion is developed. Variations of the general method are considered in which the geometry of the CCD imager's photosen-sors are varied. A CCD imager with square, rectangular, and triangular shaped pixels is considered. To compare the different methods the amount of distortion introduced to the log-spiral representation is determined by direct calculation for each of the camera configurations. The results suggest that a camera with varied pixel shapes could form a log-spiral representation with less distortion.

NASA and Texas Instruments have developed an image processor, the Programmable Remapper 1, for certain functions in machine vision. The Remapper performs a highly arbitrary geometric warping of an image at video rate. It might ultimately be shrunk to a size and cost that could allow its use in a low-vision prosthesis. We have developed coordinate warpings for retinitis pigmentosa (tunnel vision) and for maculapathy (loss of central field) that are intended to make best use of the patient's remaining viable retina. The rationales and mathematics are presented for some warpings that we will try in clinical studies using the Remapper's prototype. (Recorded video imagery was shown at the conference for the maculapathy remapping.

A system approach to correlation-based rotation and scale invariant pattern recognition is described. Familiar operations such as the coherent optical Fourier transform and noncoherent correlation are included in a concept which is strongly influenced by the need for practical manufacturing techniques. There is merit in amplitude-only filtering for noncoherent correlation, and matched filtering diffraction effects are not severe. Emphasis for innovation is on electronic subsystems. A digital remapper and analog correlation field processor are discussed. Analysis leading to a prediction of several thousand frames per second is summarized.

The previous implementations of our Epipolar-Plane Image Analysis depth measuring technique for a moving vehicle [3] demonstrated the feasibility and benefits of the approach, but were carried out for restricted camera geometries. The question of more general geometries made utility for autonomous navigation uncertain. We have developed a generalization of the analysis that a) enables varying view direction (including varying over time), b) provides three-dimensional connectivity information for building coherent spatial descriptions of observed objects, and c) operates sequentially, allowing initiation and refinement of scene feature estimates while the sensor is in motion. To implement this generalization it was necessary to develop an explicit description of the evolution of images over time. We have achieved this by building a process that creates a set of two-dimensional manifolds defined at the zeros of a three-dimensional spatiotemporal Laplacian. These manifolds represent explicitly both the spatial and temporal structure of the temporally-evolving imagery, and we term them spatiotemporal surfaces. Named the Weaving Wall, the process which builds these surfaces operates over images as they arrive from the sensor, knitting together, along a parallel frontier, connected descriptions of images as they evolve over time. We describe the sequential construction of these surfaces and their use in three-dimensional scene reconstruction. The use of the process in other domains will be highlighted by showing its application in the area of medical imaging.

An important task in many vision applications is that of rapidly estimating the orientation of an object with respect to some frame of reference. Because of their speed and parallel processing capabilities, optical correlators should prove valuable in this application. This paper considers two algorithms for object orientation estimation based on optical correlations and presents some initial simulation results.

Wide-band phase-mostly filters are notably "sharp" with respect to in-plane image changes such as rotation and scale differences. Adverse consequences include the necessity of large libraries of reference views covering the range of rotations and scales of the object being recognized. In the near vicinity of an exact match between reference and object, the phase-mostly correlation stands out against background far more strongly than necessary for object recognition. One can move that energy off the peak and into the correlation plane vicinity of the peak with appropriate adjustments to the filter. We present an algorithm that blunts a filter against scale and rotation mismatch: a blurring expressed as a gaussian convolution in the log-polar system. An expression is derived for a four-dimensional cartesian resampling kernel, and an approximate method of implementing it is given. The filter process is shown in computer simulation for frame-grabbed imagery.

Ternary correlation filters, which encode modulations of -1, 0 and +1, may be viewed as a logical step beyond binary phase-only filters. Both formulations are motivated by the prospect of relatively simple real-time implementation compared to full complex matched filters. The zero-modulation state of ternary filters affords additional flexibility and control in filter design. In particular both enhanced nontarget discrimination and reduced distortion sensitivity, relative to simple binary phase-only filters, can be achieved simultaneously as demonstrated by correlation simulations reported here.

The data processing problems associated with imaging spectrometer data are reviewed; new algorithms and optical processing solutions are advanced for this computationally intensive application. Optical decision net, directed graph, and neural net solutions are considered. Decision nets and mineral element determination of non-mixture data are emphasized here. A new Fisher/minimum-variance clustering algorithm is advanced, initialization using minimum-variance clustering is found to be preferred and fast. Tests on a 500 class problem show the excellent performance of our algorithm.

A joint Fourier transform image correlator (JFTC) that can recover the correlation signals from an undersampled interference intensity displayed on an electrically addressed spatial light modulator (SLM) at the Fourier plane is described. The implementation of a space invariant JFTC using an electrically addressed SLM at the Fourier plane requires fine sampling of the Fourier transforms' interference intensity to resolve the fringe pattern. However, with the limited space-bandwidth product (SBWP) of the SLM, only the low spatial frequencies of the input spectrum can be recorded and displayed. This can result in the loss of the fine details of the input images which can degrade the performance of the system. Here, we show that given a defined field that contains the correlation functions, the correlation signals can be formed without fine sampling of the spatial carrier frequency term and without significant losses in the fine details of the input images. The correlation functions can be recovered from the undersampled interference intensity according to the criterion for sampling of bandpass signals given that the field that contains the correlation signals is known. Space invariant multiple object detection is possible within the defined correlation field. The proposed technique is suitable for bipolar JFTC that have one-pixel auto-correlation bandwidth and very small cross-correlation sidelobes.

Information capacity of holography (Ih) is deduced from the point of view of 3-D reconstructed image, the pixels being now 3-D "Airy ellipsoids". Then Ihis compared with the information capacity of photography (I p), one of the conclusions being that sometimes Ih is less than Ip .

DNA sequence analysis has been demonstrated with optical pattern recognition techniques. New methods to optically study features of the DNA molecular code have been developed by creating new DNA sequence representations. This research involves representing DNA sequences by characters which have been designed so that their Fourier transform properties can be used to perform optical searches for non-specific sequence features. To aid in the design of these characters, a computer simulation of the optical process was developed. Matched spatial filters (MSF) were made of important DNA features using the new DNA representations and searches performed on DNA sequences. The search results were obtained using optical correlation and studied with the aid of image processing capabilities on a microcomputer. Topics to be discussed are DNA features and organization, character design, and optical pattern recognition.

The innermost layer of the cornea consists of a single layer of cells which are responsible for the active transport of fluid from the cornea to the aqueous humor. This cellular layer is about 5 pm thick and consist of cells of various shapes (pentagons, hexagons, hexagons, and octagons) although hexagons predominate. With aging and disease states, the cells become enlarged and the percent of hexagons decrease with the appearance of many five and seven side cells. The shape of the endothelial cells in the living cornea is usually evaluated from specular photomicrographs made in specularly reflected light which shows the cell borders in a square millimeter of corneal endothelium. The cell borders are traced out and the cells are analyzed with a digitizer pad, stylus, and computer software. This yields the cell area, density, coefficient of variation, and a histogram of n-sided cells. As an alternative approach to evaluate the cell shape and pattern optically, the following system was employed. The Fourier transform of the cell pattern was obtained on a digital image processor. This technique results in a fingerprint of the original cellular pattern which can be used to characterize the cellular arrangement.

The use of optical correlators for the measurement of an object's pose presents somewhat different problems than the classic object recognition problem. For example, the correlation amplitude is dependent on the in-plane rotation of 'the input object relative to the matched filter (an even more prominent condition with phase-only filters). In pattern recognition, it is desirable to construct a composite filter which is invariant to rotation so that the object will be identified no matter what its input rotation value is. On the other hand, the rotation may be one of the parameters to be determined during a docking scenario. Juday and Monroe have proposed to construct synthetic estimation filters (SEF's) designed to reduce the number of required filters, and also allow interpolations for the positional parameters between the views from which the filters were made. Preliminary work considered only in-plane rotation and used high contrast images. In this paper we report constructing SEF's for an out of plane rotation (yaw) and the consequences of using more realistic images.

The phase-only (POF) and binary phase-only (BPOF) filter have been shown to have high optical efficiency, good target discrimination, and virtually no sidelobes. We present a simple method of encoding phase information in an amplitude mode that has the same advantages of a POF or BPOF so that amplitude modulating SLMs can be used.

Binary phase-only filters (BPOF) are a viable candidate for the replacement of matched filters in real-time image processing and pattern recognition applications. The original BPOFs were binarized versions of the real or imaginary parts of the Fourier transform. Recently, a filter has been proposed which is the binarized Hartley transform. In this paper, the noise performance of the binarized Hartley phase-only filter is analyzed and compared to other BPOFs as well as the phase-only and matched filter. It is well known that the matched filter optimizes signal-to-noise ratio. Further, it can be shown that the matched filter phase is the optimum phase-only filter. A bound on the BPOF signal-to-noise ratio relative to that of the phase-only filter is derived. Finally, the analysis is applied to the generalized form of the binarized Hartley transform.

A technique for detecting objects in noisey range imagery and within a cluttered background is discussed and demonstrated on real data. The approach is based on a simple region growing technique which uses the range difference between neighboring pixels in the image as the similarity measure in the growing process. A region list is generated by the process which includes several region properties. Some of these properties are used to validate regions as detections based on a priori knowledge of general object structure. The performance of the techique is tested at several spacial resolutions and with ambiguous range data. Results of varying the similarity measure threshold is also shown. An approach to object orientation estimation is also discussed.

In order to increase the performance of a laser range sensor system one can develop techniques for detecting and classifying objects with undersampled imagery. The goal is to be able to detect and classify a limited set of objects with less than 100 samples on the object. The objects of interest can be modeled as rectangular solids. The sensor system discussed has a low depression angle scan geometry (10 to 20 degrees below horizontal); this means that the sensor sees both the front face and top of the object. The sensor undersamples in the downtrack direction with sample spacing approximately half the width of the objects of interest. The crosstrack sample spacing is close to the Nyquist criteria. This paper assumes that a detection window containing the object of interest is available. Three techniques for extracting the basic geometric features (i.e. length, width and height) are discussed. The first two approaches to extracting the length and width treat the object as a blob and use the detected extrema. These techniques will be compared to another which detects at least two opposing corners and the orientation of the object.

A method for classifying objects invariant to position, rotation, or scale is presented. Objects to be classified were multifunction laser radar data of tanks and trucks at various aspect angles. A segmented doppler image was used to mask the range image into candidate targets. Each target is then compared to stored templates representing the different classes. The template and the image were transformed into the magnitude of the Fourier transform with log radial and angle axis, lF (Ln r , 0)1, feature space. The classification is accomplished using the shape of the correlation peak of the IF (Ln r , 0)1 planes of an image and a template. A neural network was used to perform the classification with a classification accuracy near 100%. The neural network used in this study was a multilayer perception using a back propagation algorithm.

An approach to the design of associative memories and pattern recognition systems which utilizes efficiently hybrid architectures is illustrated. By associative memory we mean a database organization that supports retrieval by content and not only by name (or address), as is the case with practically all existing database systems. The approach is based on a general, metric, model for pattern recognition which was developed to unify in a single model two basic approaches to pattern recognition-geometric and structural-preserving the advantages of each one. The metric model offers the designer a complete freedom in the choice of both the object representation and the dissimilarity measure, and at the same time provides a single analytical framework for combining several object representations in a very efficient recognition scheme. It is our fervent hope that the paper will attract researchers interested in the development of associative memories or image recognition systems to experiment with various optical dissimilarity measures (between two images) the need for which becomes so acute with the realization of the possibilities offered by the metric model.

An essential feature of a practical automatic image recognition system is the ability to tolerate certain types of variations within images. The recognition of images subject to intrinsic variations can be treated as a sorting task in which an image is identified as a member of some class of images. Herein, the maximum-likelihood strategy, an important tool in the field of statistical decision theory, is applied to the image classification problem. We show that the strategy can be implemented in a standard image correlation system and that excellent classification results can be obtained.

In this paper a system for representing and recognizing three dimensional (3-D) objects is introduced. This system depends on segmenting object surface into surface patches, each patch belongs to a single type of surface primitive such as plane, sphere,...,etc. The concept of " primitive classifier " using discriminant or decision function is introduced. This classifier is designed according to what will be called "prediction functions" for each type of primitive surface. The 3-D object is represented by a model containing selected features obtained from its surface patches. According to this type of representation, a recognition algorithm for 3-D objects is designed. The performance of these techniques is tested using computer simulation.

In this note we review the object and relational c-means algorithms, and the theory asserting their duality in case the relational data corresponds to an inner-product induced measure of distance between each pair of corresponding object data. Past numerical results are given here along with new extensions in order to study the effect of the choice of pairwise distance measure on the relational partition obtained.

If we superficially compare the computer pattern treatment with the functioning of the brain, the advantages are evident. The brain uses the retinal images accurately and without difficulty, and gives a swift response, even though at the same time it must take into account information received from our surroundings and our memory.

In this paper, the use of Cylindrical Multivalued Transform (CMLT) for defining an Integral Transform Measure is examined. This Measure is robust to the effect of noise and quantization errors. These signatures are used as low level primitives in computer vision system for the purpose of recognizing curves and surfaces (shapes). Algorithms for evaluation of these signatures are presented. Properties of these signatures are also discussed.

Various range sensors are being used as part of a machine vision system to capture the surface of a 3D object. Significant structural changes on the surface, such as zero gaussian curvature contours, can be detected and regarded as 3D curves using differential geometry tools. In this paper, we present a technique for recognition of partially visible 3D curves which can be employed to identify 3D objects. A local/global represenation of a 3D curve is obtained by decomposing it into a number of invariant 3D curve segments. Based on curvature and torsion functions, feature vectors are assigned to the curve segments. A distance measure is then defined between two decomposed 3D curves using their corresponding feature vectors. The recognition of an occluded curve among known reference curves is achieved by obtaining the smallest distance measure. Experimental results on both simulated and real data are presented.

A real-time, Monte Carlo method for computation of image moments is presented. Input images are reduced to very low light levels, and imaged onto a two-dimensional, position-sensistve photon-counting detector. The coordinates of the detected photoevents are independent random variates, which are used to produce Monte Carlo estimates for the momentinvariants of the input image. Experimental results demonstrate that 5000 detected photoevents provide accurate estimates for several moment invariants of a typical input image.

In this paper, a new method of the group direction difference chain code representation is presented which can be used to represent the curves or shape borders, and also can be used for the features extraction of shape borders. The idea of this method is to construct a chain codes based on the Freeman chain codes whose values represent the average changes in direction along the boundary of an any shape. For the sake of overcoming the noise in Freeman chain codes, in this paper the concept of direction smoothing is proposed, and a transform from Freeman chain codes to the direction smoothing chain codes is developed. Finally, the method and algorithm for obtaining the group direction difference chain codes are presented.

Spherical harmonics are introduced as a modeling tool. Their properties suggest that they are a representation well suited to object recognition. We explain how models are instantiated from inputted data sets and introduce a natural coordinate frame in which to build these models. We address the question of whether an object should be represented as a single model or be described rather as a collection of parts. As regards the latter, we define our understanding of what, in effect, constitutes a part. Finally we discuss the application of this representation to the tasks of object recognition and to scene rendering.

A sequence of structuring elements S = {S1,...,SN} is said to be increasing if it has the property that for each i, Si+i ⊃ Si. In general, such sequences are made up of elements with similar shapes but different sizes; e. g., lines, squares, octagons, and disks. A morphological operation ψ is said to be monotonic with respect to an Increasing structuring element sequence S, if, for any set X, either: (X 'F Si+1) 2 (X 'F Si), Vi (Monotonic increasing) Or (X AIF Si) 3 (X III Si+1), Vi (Monotonic decreasing) Dilation is monotonic increasing while erosion is monotonic decreasing. These properties make it possible to unambiguously classify every pixel in a binary image by associating each with one of the elements in the sequence S. Morphological openings and closings are also monotonic, but only if an additional property holds for the sequence 5, namely, that for each i, there exists a structuring element T such that Si+1 = (Si ⊕ T), or in other words, Si and Si+1 must be similar in shape up to a dilation. In the digital world, squares, hexagons, and octagons are similar but digital approximations to disks are not. This poses problems for trying to generate morphological shape and size distributions based on very accurate digital disks. This paper proves the monotonicity properties for erosions, dilations, openings, and closings, and shows how pixel distributions or classifications based on shape can be generated from these properties. It also discusses the problem posed by digital disks, and describes one method of circumventing it.

A model-based vision system requires models in order to predict object appearances. How an object appears in the image is the result of interaction between the object properties and the sensor characteristics. Thus in model-based vision, we ought to model the sensor as well as the object. Previously, the sensor model was not used in model-based vision or, at least, was contained in the object model implicitly. This paper presents a framework between an object model and the object's appearances. We consider two aspects of sensor characteristics: sensor detectability and sensor reliability. Sensor detectability specifies what kinds of features can be detected and in what condition the features are detected; sensor reliability is a confidence for the detected features. We define the configuration space to represent sensor characteristics. We propose a representation method for sensor detectability and reliability in the configuration space. Finally, we investigate how to apply the sensor model to a model-based vision system, in particular, automatic generation of an object recognition program from a given model.

Given a CAD model of an object and a set of inspection specifications, we would like to automatically generate the vision procedure to inspect a part that is an instance of the model. Since the position and orientation of the part may be wholly or partially unknown, the first step in the procedure is to determine the pose of the object. Assuming the sensor involved is a CCD camera, this reduces to matching the features extracted from a two-dimensional graytone perspective projection image of the object to the corresponding three-dimensional features of the model. Since 2D to 3D matching is more complex and time consuming than 2D to 2D matching, our preference is to match a data structure representing features and their spatial relationships extracted from the image to a similar 2D data structure generated from the CAD model. Our approach is to use the CAD model to predict the features that will appear in different views of the object under different lightings and use these visible features to generate a set of view classes for use in the matching. A view class is a cluster of views of the object which all produce similar data structures. Then a single representative data structure can represent the entire cluster of views and be used to match to the structure extracted from the image. Important questions that must be answered are 1) how do we predict features from CAD models without generating entire artificial images of the object, 2) what is a good representation for the features extracted from one view, 3) what criteria should be used for forming view classes, and 4) how should the matching from part structure to view class representatives be achieved most efficiently. In this paper we will report on our ongoing research in these areas. Keywords: matching, view class, CAD model, relational pyramid This research was supported by the National Aeronautics and Space Administration (NASA) through a subcontract from Machine Vision International and by Boeing Commercial Aircraft Company.

A knowledge-based 3D object recognition system has been developed. The system uses the hierarchical structural, geometrical and relational knowledge in matching the 3D object models to the image data through pre-defined primitives. The primitives, we have selected, to begin with, are 3D boxes, cylinders, and spheres. These primitives as viewed from different angles covering complete 3D rotation range are stored in a "Primitive-Viewing Knowledge-Base" in form of hierarchical structural and relational graphs. The knowledge-based system then hypothesizes about the viewing angle and decomposes the segmented image data into valid primitives. A rough 3D structural and relational description is made on the basis of recognized 3D primitives. This description is now used in the detailed high-level frame-based structural and relational matching. The system has several expert and knowledge-based systems working in both stand-alone and cooperative modes to provide multi-level processing. This multi-level processing utilizes both bottom-up (data-driven) and top-down (model-driven) approaches in order to acquire sufficient knowledge to accept or reject any hypothesis for matching or recognizing the objects in the given image.

Three-dimensional model-based computer vision uses geometric models of objects and sensed data to recognize objects in a scene. Likewise, Computer Aided Geometric Design (CAGD) systems are used to interactively generate three-dimensional models during the design process. Despite this similarity, there has been a dichotomy between these fields. Recently, the unification of CAGD and vision systems has become the focus of research in the context of manufacturing automation. This paper explores the connection between CAGD and computer vision. A method for the automatic generation of recognition strategies based on the geometric properties of shape has been devised and implemented. This uses a novel technique developed for quantifying the following properties of features which compose models used in computer vision: robustness, completeness, consistency, cost, and uniqueness. By utilizing this information, the automatic synthesis of a specialized recognition scheme, called a Strategy Tree, is accomplished. Strategy Trees describe, in a systematic and robust manner, the search process used for recognition and localization of particular objects in the given scene. They consist of selected featureS which satisfy system constraints and Corroborating Evidence Subtrees which are used in the formation of hypotheses. Verification techniques, used to substantiate or refute these hypotheses, are explored.

The marriage of machine vision systems and CAD databases will be useful in solving many industrial problems. In this paper, we discuss our approach to recognizing objects in range images using CAD databases. The models in the databases are used to generate recognition strategies and to test hypothesis generated by the early recognition system. The role of features has been recognized since early days of pattern recognition research. Here we discuss what are the desirable characteristics of features in a 3-D object recognition system. We formally define the features used in our approach and discuss their strengths and weaknesses.

This paper presents a model-based object recognition method which combines a bottom-up evidence accumulation process and a top-down hypothesis verification process. The hypothesize-and-test paradigm is fundamental in model-based vision. However, research issues remain on how the bottom-up process gathers pieces of evidence and when the top-down process should take the lead. To accumulate pieces of evidence, we use a configuration space whose points represent a configuration of an object (ie. position and orientation of an object in an image). If a feature is found which matches a part of an object model, the configuration space is updated to reflect the possible configurations of the object. A region in the configuration space where multiple pieces of evidence from such feature-part matches overlap suggests a high probability that the object exists in an image with a configuration in that region. The cost of the bottom-up process to further accumulate evidence for localization, and that of the top-down process to recognize the object by verification, are compared by considering the size of the search region and the probability of success of verification. If the cost of the top-down process becomes lower, hypotheses are generated and their verification processes are started. The first version of the recognition program has been written and applied to the recognition of a jet airplane in synthetic aperture radar (SAR) images. In creating a model of an object, we have used a SAR simulator as a sensor model, so that we can predict those object features which are reliably detectable by the sensors. The program is being tested with simulated SAR images, and shows promising performance.

Few three-dimensional object representation systems allow non-rigid connections between components of the object. We define a representation scheme that permits parameterized non-rigid connections, allowing one definition to specify a range of permissible configurations of an aggregate object. This representation can be used to generate 3-D instantiations of particular configurations, and 2-D projected images of particular 3-D instantiations. General issues involved in constructing such an object representation are outlined. The syntax of component description and connection types for our specific system is reviewed, along with the semantics of the allowable ranges of movement associated with each connection type. The actual representation system modules are also described.

Parallel distributed, connectionist, neural networks present powerful computational metaphors for diverse applications ranging from machine perception to artificial intelligence [1-3,6]. Historically, such systems have been appealing for their ability to perform self-organization and learning[7, 8, 11]. However, while simple systems of this type can perform interesting tasks, results from such systems perform little better than existing template matchers in some real world applications [9,10]. The definition of a more complex structure made from simple units can be used to enhance performance of these models [4, 5], but the addition of extra complexity raises representational issues. This paper reports on attempts to code information and features which have classically been useful to shape analysis into a neural network system.

We present a recipe for preparation of holograms without use of photography or any special equipment. The holograms are computer-generated on a personal computer, the data is printed on a Hewlett Packard jet printer, and the hologram is printed on an overhead transparency in the last of three photo-copy reductions. The image can then be reconstructed with a simple laser, collimator, lens, vidicon camera, set-up. The presentation includes a hologram which a reader can photo-copy and use directly.