In the Optical Character Reader (OCR) system design, the character segmentation technique is important. For example, the Automatic Mail Address Reader is required to manage printed characters of many font types and poor print quality. In this case, OCR performance will be affected by character segmentation technique. This paper describes two new methods for character segmentation under more general conditions. The character segmentation problem can be formulated and classified as a pitch estimation problem and a character sectioning decision problem. These problems are resolved by using a statistical analysis method based on least square error function and a dynamic programing method with the minimum variance for separation between candidate positions in a line image. The effectiveness of the proposed methods has been evaluated through actual mail address segmentation experiments.

In the automatic recognition of tactical targets in FLIR images, it is desired to obtain an accurate and precise representation of the boundary of the targets. It is very important since the features used in the classification of the target are normally based on the shape and gray scale of the segmented target and therefore the performance of a statistical or a structural classifier critically depends on the results of segmentation. Generally, only the gray scale of the image is used to extract the target from the background. The segmentation thus obtained normally depends upon several parameters of the technique used. It is possible to obtain better segmentation by using other sources of information present in the image such as contextual cues, temporal cues, gradient, a priori information etc. In this paper we consider specifically the use of gray scale together with the edge information present in the image to obtain more precise segmentation of the target than obtained by using gray scale or edge information alone. A model of FLIR images based on gray scale and edge information is incorporated in a gradient relaxation technique which explicitly maximizes a criterion function based on the inconsistency and ambiguity of classification of pixels with respect to its neighbors. Four variations of the basic relaxation technique are considered which provide automatic selection of threshold to segment FLIR images. A comparison of these methods is discussed.

A two-level feature extraction classifier using a geometrical-moment feature space is described for multi-class distortion-invariant pattern recognition. The first-level classifier provides object class and aspect estimates using multi-class Fisher projections and optimized two-class Fisher projections in a hierarchical classifier. Aspect estimates are provided from ratios of the computed moments. The second-level classifier provides the final class estimate, distortion parameter estimates and the confidence of the estimates. Extensive test results on a ship image database are presented.

Pyramidal structures created from resolution sequences of imagery contain a hier-archical structure induced by the resolution function. Conversion of this pyramidal structure into a Resolution Syntax Tree (RST) lends itself to a hierarchical recognition process. The conversion is accomplished by creating a representation of the global and local connectivity of the internal structure segmentation at each level of the pyramid. The recognition process uniquely classifies targets by determining a structural similarity with a knowledge base. The analysis is a matching strategy that has its foundation in formal languages using a graph, theoretic approach. A stochastic metric is added to the recognition process to account for real-world variations. The recognition process begins by down-selecting the knowledge base into a working set using a priori knowledge of the image viewing geometry. The hierarchical recognition individually tests each entry in the working set against the target to determine a probability of structural similarity. The process begins at the root of the RST, which corresponds to the global information of the pyramidal structure. The segmentation of the internal structure produces a graph; the edges of the graph represent the boundaries found during segmentation, and the vertices are the endpoints of these boundaries. A structural match between a working set entry and the target is obtained by determining a subgraph common to both boundary graphs. The subgraph is found by forming a match in the bipartite graph developed from two sets of boundary graph vertices. A stochastic metric is derived based on the commonality of the subgraph to the two boundary graphs. The recognition process may be recursively applied to the next level of the RST, utilizing the level-to-level connectivity. Recursion is performed only as warranted by the stochastic metric determined at the previous level.

The recognition of targets in thermal imagery has been a problem exhaustively analyzed in its current localized dimension. This paper discusses the application of artificial intelligence (AI) technology to automatic target recognition, a concept capable of expanding current ATR efforts into a new globalized dimension. Deficiencies of current automatic target recognition systems are reviewed in terms of system shortcomings. Areas of artificial intelligence which show the most promise in improving ATR performance are analyzed, and a timeline is formed in light of how near (as well as far) term artificial intelligence applications may exist. Current research in the area of high level expert vision systems is reviewed and the possible utilization of artificial intelligence architectures to improve low level image processing functions is also discussed. Additional application areas of relevance to solving the problem of automatic target recognition utilizing both high and low level processing are also explored.

A set of uniform magnitude, minimal-order moment invariants is introduced. The relationship to the Hu and other invariants is explained. It is shown how the present invariants circumvent some of the more serious limitations of the Hu set. For example, it is shown that the Hu invariants are algebraically dependent and how this arises. Graphic representations in the form of hidden-line surfaces in feature space are presented. A proposed basis for "aspect-independents" is outlined.

A new approach to the problem of classifying surface materials in satellite multi-spectral imagery is described and demonstrated in this paper. Surface material classes are defined heuristically using rules which describe the typical appearance of the material under specified conditions in terms of relative image measures. A knowledge-based approach allows expert knowledge of the domain to be used directly to develop classification rules. An expert system is currently being developed in the Zetalisp/Flavors programming environment on the Symbolics 3600 Lisp Machine. An example of its use in classifying Landsat Thematic Mapper imagery is presented.

This method relates to the inspection of pharmaceutical filled capsules. The object of this inspection method is to detect mainly structural defects of the capsule with or without printing. To recognize the characters which are printed on the surface of the capsule is not an object. Capsules consist of a cap and body which are telescopically fitted together, and the bodies are filled with medicinal material. The two spherical ends and near the center region of each capsule should be inspected for defects. Further the area where printing may cover should be inspected for indications of such defects as a foreign material, a body dent or a speck. On the basis of projection technique, the body/cap boundary points can be distinguished from the printing edge points near the body/cap boundary, and the size of the area which the printing may cover can also be computed.

For analysis of two dimensional plastic deformation of coarse crystal aluminum sheet specimen under several forces, a new microcomputer-aided analytic method to monochromatic video images of plastic deformation of fine square grids which have excellent contrast to the specimen and good deformability with the specimen was developed. Images on the specimen in testing are taken continuously with a CCTV camera having an object lens with suitable magnification and recorded into a video tape as usual. Selected images are converted into digital data and written into a floppy disk one and after another, and an image made by binary data is transfered monochromatically on a colored, high discrimination display of a micro-computer, and Cartesian coordinates of each cross point after a new fining processing are read out, several deformation distribution map and strain map in various cases shown by vectorial expression is gotten on a display and a printer also. The analytic process, an experimental instance and some points at issue are reported here.

Machine vision systems have undergone dramatic changes in the last few years due to the rapid advances in microprocessors and specialized digital signal/image processing electronics. (1,2) This paper is a report of a alignment, inspection and measurement system, KLAASP, (KLA Automation System for Process Control). The applications of this machine vision system are initially directed toward the semiconductor test and assembly area where extensive optical alignment and inspection are required. (3,4)

To keep pace with the trend towards increased circuit integration, printed circuit patterns are becoming denser and more complex. A variety of automated visual inspection methods to detect circuit defects during manufacturing have been proposed. This paper describes a method that is a synthesis of the reference-comparison and the generic property approaches that exploits their respective strengths and over-comes their respective weaknesses. It is based on the observation that the local geometric and global topological correctness of a printed circuit can be inferred from the correctness of simplified, skeletal versions of the circuit in a test image. These operations can be realized using simple processing elements which are well suited for implementation in hardware.

Automated visual inspection of printed wiring is providing an opportunity for research in completely automated optical systems technology. This paper addresses device design considerations particularly with regard to the important problem of turning an optical inspection device into a manufacturing process analysis and control tool.

We discuss the design of a large scale Content Addressable Array Parallel Processor (CAAPP) for low, medium and high level vision processing. This new architecture combines associative processing with global broadcast and response to and from an array of cells, and array processing via local cellular square neighborhood computation. The capabilities of the CAAPP allow us to close the feedback loop between high level processing and low level processing by supporting communication between different representations of an image. The CAAPP would provide a means of mapping the signal level (iconic) pixel-based representation of an image into a symbolic intermediate level representation suitable for high level vision processing.

In the past several years, design and implementation of an acceptable user interface has become the major task in developing an image processing software system. The general image processing algorithms required to address a wide range of applications have been developed and well documented in recent years. The real challenge in software system design has become the development of user interfaces applicable to a particular applications or usage scenario.

This paper describes the concept and implementation of a Software Digital Image Processing Calculator (SDIPC). Many of today's digital image processing facilities either possess or have remote access to high-speed floating-point array processors. Unfortunately, these processors are often underutilized due to the difficulty in programming them. The SDIPC alleviates this problem by allowing the user to specify digital image processing algorithms using a reverse-polish expression syntax. Most common arithmetic and transcendental functions may be specified in such an expression. The user's expression is mapped into a library of mathematical functions which are then executed in the proper sequence in a floating-point array processor. Because of its generality, the performance of the SDIPC is slightly less than that which could be obtained by an application program customized to execute a specific algorithm. This is more than compensated for by increased user productivity, flexibility, and creativity. Several examples of applications and their performance are described, including matrix transformations and gradient magnitude and direction computation.

This paper describes a hardware module to be a part of a general image computer that is being developed in our laboratory. The design of this module was started to provide us with a powerful processor suited for several classes of image filtering operations on a single board at a very high speed and also to gain insight and experience in hardware that is suitable for adaptive filtering of images. A prototype of the module is now being built.

An automatic stress analyzer from photoelastic fringes using personal computer is developed. The photoelastic fringes are taken in by TV camera. On the basis of the image processing, the isochromatic fringes are extracted, and the fringe orders are determined. The principal stresses are separated with the combination of the sum of principal stresses calculated by the finite element analysis of Laplace's equation and the difference of principal stresses obtained experimentally from the isochromaic fringes. The results obtained are put out to monitor TV, printer, plotter and disk unit. The analyzer developed is low-price, easily operated, and suits to analyze stresses within the whole domain under consideration.

Presented here is a simple, easily implemented method for an FFT-based implementation of image cross-correlation which provides for the proper normalization required to obtain useful, easily interpreted output. The results are applied to character recognition. The algorithm is structured such that the data is processed in a highly parallel fashion in order to facilitate a fast, efficient hardware implementation. A discussion is made of the relative merits of using this method.

The problem of image segmentation is approached via finite state machines. Finite state machines are excellent tools in processing various vision tasks. In this paper, segmentation of radiograph images of welding scenes are considered. A Mealy machine has been designed to remove background nonhomogeneity and a Moore machine has been designed to isolate the welding defects from the background. Result of segmentation of several radiograph images using these finite state machines are given.

The Pipelined Resampling Processor (PRP) was presented to this conference in 1983 in concept as a practical solution to real-time high resolution geometric image rectification needs. In the present report we present performance results obtained from the PRP development unit completed in late 1983. The small size, weight, and power requirements of the PRP and its high throughput make it very well suited for space and airborne applications where goemetric correction of image data must be done autonomously in real-time. This high resolution geometric correction is a necessary adjunct to applications using frame differencing or frame averaging for motion compensation, moving target indication, noise suppression or data compression as well as applications requiring precise correction of focal plane sampling distortions. By equipping the PRP with manual controls, image geometry can be manipulated at video rates from a console to achieve a much higher image analysis throughput than is possible for more general purpose processing facilities. In mapping, merging, classifi-cation and registration applications, interactive video rate processing will be important in bringing these and other image analysis techniques out of the laboratory and into an operational environment.

This paper deals with the development of a general architecture for image processing after considering important aspects of image processing algorithms commonly used today. The algorithms are linked by mutual operations and using this fact leads to an architecture suitable for a wide variety of operations. An architecture will be developed that spans a wide range of functionality through systematic expansion, making it fit for low cost production and commercial consumption. It is also the intent of this paper to introduce a general theory about such systems, and to provoke further thinking along these lines.

The planar image of an object can be represented as a "codon" list of maxima, minima and zeros of the image's curvature. This representation is invariant over rotation, translation, scale, and other distortions of the image, making it a useful representation for visual recognition or inspection. Calculating this representation from a digital image is difficult because of the granularity of the image and the possibility of many scales for the object (e.g., a fractal). We present an algorithm for examining a digital image at various scales, computing curvature, and symbolically encoding this information. This algorithm is designed for a digital frame buffer and video arithmetic logic unit.

This paper describes a flexible gray scale image enhancement scheme coupled with segmentation algorithms to automatically describe elemental shapes arising in a wide variety of images of interest in machine vision applications. The enhancement algorithm is a locally adaptive Fourier filter configured so as to easily perform either contrast enhancement or additionally apply more complex Fourier filters to enhance periodic features. The enhanced images are then presented to a thresholding and region filling algorithm which breaks the objects of interest into elemental shapes. These shapes are characterized by simple measures such as size, perimeter, and Euler number, and feature extraction tasks are built on the basis of these descriptors. The method has been applied to fingerprint classification, seismic data inspection, and automated handling of packages.

Advances in imaging technology coupled with development of high speed computer architectures now support use of grey scale and color processing for real time computer vision applications. Development of a general system for 2D pattern recognition will include an analysis of the optical properties of the materials and development of sensor specifications based upon this physical level knowledge. In grey scale or color imagery it is necessary to apply sophisticated techniques to segment the scene into components. This paper describes the elements of a vision system which is used to process images of two dimensional objects having varying color or shades of grey. Examples will be shown for some 2D vision applications.

A variety of clinical research and industrial applications involve the quantitative analysis of images from the optical microscope. Digital image analysis techniques are used to measure the size, shape, density, texture and number of objects in an image. The PSICOM 327 digital imaging system is useful for a variety of quantitative image analysis applications. This paper describes the system and some of the applications for which it is suited.

Industrial morphology combines the dominant paradigms of machine vision, image processing and pattern recognition, into a cohesive framework of algebraic fundamentals and operational systems. Machine vision systems based on the principles of industrial morphology represent a logical advancement in machine vision technology, combining the flexibility and speed of special parallel architectures for image processing, high level languages for application programming, and statistical classification and learning for interactive, menu driven calibration and setup.

In general image processing and pattern recognition laboratory, CT machine is not available. So we use simulated method to research image reconstruction algorithm and develop the application software. A Fortran application software is completed. The result is quite satisfied but bring with some errors.

A technique is described for designing two-dimensional finite-length impulse response digital filters using linear programming. The principal advantage of the technique is the ability to design a filter while controlling aspects of the spatial domain and the frequency domain simultaneously. The frequency response of the filter can be smoothed, for example, by forcing the magnitude of the impulse response to decrease sequentially from the center of the filter to the edges. This smoothing technique offers the additional advantage of requiring only three directions of the frequency response to be specified. The technique has been successfully used to design image enhancement filters for a wide range of imagery. Typical image enhancement filters designed using this method are presented.

In this paper we shall introduce a new approach to interpolation, namely the adaptive median interpolation algorithm. We shall discuss applications and stress the importance of interpolation in images. In order to understand the limitations of linear interpolation schemes we will review the methods that have been used to date and their disadvantages. We shall then illustrate how the median interpolation method overcomes the previously encountered problems. We shall conclude with a real time implementation.

An algorithm is described that is capable of determining and analyzing the placement of features within the target's image using the available edge information. This approach allows for the improved placement of the target gate about the target and the ability to select a terminal homing point for endgame or terminal tracking. The algorithm is simple but apparently effective against various targets situated against nominally cluttered backgrounds and is implementable as a real-time digital process with a relatively small amount of hardware. The approach is based on the use of thresholded edge information found within the target track gate.

A set of multistage image processing algorithms have been developed to do change detection in image sequence analysis. These algorithms are tailored for the purpose of moving target identification (MTI). The image processing algorithms have been coupled with a simple tracking-detection algorithm. The resulting combination of processing shows very good performance in detecting targets in simulated mosaic IR images with projected probability of false alarm less than one per hundred billion frames and the probability of detection for targets within the model approaching unity.

A theoretical background of image and geometrical transformations is introduced. Digital signal processing techniques as applied to image processing are discussed. The problems associated with real-time image processing are presented. A distributed system employing multiprocessors to process data in parallel is described.

The potential of the analysis of variance (ANOVA) as a pattern recognition tool is evaluated and two specific examples are presented. First, we present the 2-way ANOVA as employed in NASA's Project Oasis - the design of a signal detector for the search for extraterrestrial intelligence (SETI). In the Oasis design, ANOVA is capable of detecting non-random signals in an 8 million x 1000 matrix band of noisy signal data in quasi real-time. The presence of a signal with low signal-to-noise ratio is detected by the ANOVA by a rejection of the null hypothesis for certain row and/or column effects of compacted submatrices. In a somewhat analogous fashion, a 3-way ANOVA is used to detect camouflaged moving targets. Targets not differentiable from their background by the eye can be detected by the ANOVA.

A two-step procedure is developed for restoring low light level images degraded by a linear space-invariant blur. The first step uses a linear minimum mean square point estimate of the blurred image. The second removes the blur through a constrained linear least squares technique where the error of the first step is treated as additive noise. No prior knowledge of the object is required, as the procedure utilizes sample statistics in an analysis window around each received image element to develop the estimators. The size of the window is adaptive, as it is adjusted for the next image element according to an activity index computed for the current window. In experiments with simulated photon-noise-degraded, lineal-motion-blurred images, the efficacy of the procedure is demonstrated visually and measured by signal-to-noise ratio improvements.

This paper describes an adaptive filtering algorithm that uses edge information, contained in the original image, to control a space varying convolution type filter. This adaptive filtering approach is more powerful than the traditional filtering methods because it uses some kind of knowledge about the image, whereas simple linear filtering for instance, does not use any knowledge about the image. In this case, the knowledge that is used is the location, direction and intensity of the edges. The edge information is enhanced with a relaxation process before it is used to control a filter.

This paper addresses the problem of local/spatially-variant/adaptive image processing based on direct estimates of local autocovariance functions. In order to quantify the non-stationarity of images, and often, to implement spatially-variant processing, we require estimates or measurements of the local image statistics, specifically the autocovariance function. The simplest way to achieve this is to divide the image into N x N - pixel sub-blocks (e.g. N = 16), and calculate the usual biased or unbiased autocovariance function of each sub ock. In effect, each subblock is treated as part of a wide-sense stationary field. It is well-known, however, that reliable power spectral estimates require much larger amounts of data. Nevertheless, as our work shows, it is possible to obtain useful maps of local autoco-variance parameters if we assume simple parametric autocovariance models. Specifically, we employ popular first-order models, such as the nonseparable exponential model. We discuss a procedure for estimating local autocovariance parameters. The resulting parameters are seen to correlate with observed signal activity. We also outline techniques for spatially-variant image processing - coding, restoration, and enhancement - based on local statistics. Processed examples are given.

In the registration of two dimensional images that are obtained from two different sensors and from geometrically different locations there are distortions. These distortions can cause errors in the navigational systems that use them. In this paper a lower bound on the registration error is derived by using a model that treats the sensed and reference images as the two ends of a communication channel. Then rate distortion theory is applied to obtain a relationship between the registration error and the statistical properties of the images. The resulting bound can be used to predict the registration error, in the selection of the proper reference images, and in comparing the performance of the image registration systems.

Two iterative algorithms are presented for phase retrieval from intensity measurements for two-dimensional complex-valued sequences. The first algorithm constructs a linear combination of present and past iteration outputs using a relaxation parameter. The second algorithm utilizes a Kalman-type filter to combine estimates of object domain magnitude. A new error metric is defined that uses residual monitoring to show the relationship of the error to the reconstructed image quality. These algorithms are first tested against real-valued images to investigate their convergence chacacteristics. Both algorithms are then applied to infrared images which are characterized by low-contrast and noise-corruption. The question of uniqueness is also addressed for two-dimensional complex-valued sequences.

Dither rendition of images is a well-known method to produce pseudo-gray scale images using bi-level output devices. Moreover, it reduces the data to 1 bit/pixel. For fixed sub-picture (matrix) size there is, however, a contradiction between resolution and the dynamic range of the gray scale. In this paper we show how this conflict may be overcome using variable resolution. For this purpose we introduce a detail-dependent segmentation of the image into variable-size subpictures. We describe first the construction of suitable dither matrices of size 2mx2m, with m = 2,3,4. The next section deals with adaptation criteria to obtain the appropriate pseudo-pixel size. A major aspect of the paper is the fast in-place computer method to break the full picture up into subpictures of appropriate size. This sub-picture size depends adaptively on the local "activity" of the picture. The method is demonstrated using 8-bit test pictures and shows good gray scale rendition and sufficient de-tail resolution.

An algorithm for the detection of the boundaries of coronary arteries in coronary angiograms is presented. The algorithm constructs a parametric model of the film density along the lines perpendicular to the vessel image and estimates the parameters at each perpendicular line along the vessel, based on the observed density values. The parameters of the model include the vessel diameter and centerpoint, and account for the structure of the background as well as the distortions introduced by the imaging system. The spatial continuity of the vessel is also incorporated into the model and significantly improves the accuracy of the estimation procedure. The algorithm has been tested on synthetic data, on x-rays of contrast-medium-filled cylindrical phantoms obtained over a wide range of radiographic conditions, and on real coronary angiograms. This algorithm has better performance than methods which find the points of maximum slope of the film density at each perpendicular line.

Erosion and dilation of images were compared with other edge detection techniques on a variety of marine organisms. Under certain conditions the erosion and dilation technique gave better results. The critical problem resolved by our approach was low contrast imaging of randomly oriented objects that displayed random variations due to appendages that frequently appearred with marine biological samples. A multicomputer system was developed to perform image processing and morphological feature extraction on large number of samples. Emphasis was given to system reliability and expandability, allowing for performance at a reduced rate when one or more computers malfunctioned. The system currently operates with seven computers but can be expanded to contain up to seventeen. Classification accuracy on zooplankton samples from New England coastal waters was approximately 92%.

The majority of research in the area of image analysis over the last several years has centered on ground-based object and region analysis. Recent events have stirred an interest in the detection and classification of aircraft in flight. This paper surveys the six algorithms which have been successful applied to the problem of aircraft classification. Each algorithm is analyzed in terms of relative strengths and weaknesses. Summary results of the filter operators, aircraft types, evaluation imagery, problems addressed, and algorithm assumptions are presented for each approach considered.

The use of operating curves for characterization and evaluation of automatic target recognizers (ATRs) is presented. A unifying mathematical framework is developed which inter-relates individual ATR metrics, shooting scores, high level weapon systems metrics and operating curves. A methodology is described which allows operating curves to serve as a key communications medium between ATR designers and weapon systems analysts.

Light-microscopic images of three-dimensional objects are distorted by defocused projections into the focus plane. A reconstruction procedure for three-dimensional light-microscopic images based on the linear system theory is described. Reconstruction is executed by inverse filtering of the image using an approximation of the three-dimen-sional optical transfer function of the microscope. The derivation and the features of this transfer function are out-lined. The reconstruction procedure permits to achieve a significant improvement of the image quality. The recon-struction result of a test object is shown.

Band-to-band coregistration of multispectral remote sensing images can be achieved by electronic signal processing techniques rather than by costly and difficult mechanical alignment. This paper describes the results of a study of the end-to-end performance of electronic registration. The software simulation includes steps which model the performance of the geometric calibration process, the instrument image quality, detector performance and the effects of achieving coregistration through image resampling. The image resampling step emulates the Pipelined Resampling Processor, a real-time image resampler developed by Ball Aerospace Systems Division. The study demonstrates that the electronic alignment technique produces multispectral images which are superior to those produced by an imager whose pixel geometry is accurate to 0.1 pixel rms. The implications of this approach for future earth observation programs are discussed.

Correlation techniques are used to identify characters on integrated circuit components for the purpose of verifying that the correct component has been located on the printed circuit board prior to wave soldering. The circuit components are inspected using a computer controlled vision system and the appropriate character string located. The individual characters are isolated within the string and identified. The resultant is an aggregate number that identifies the component and is compared against a location database to verify that the component is indeed correctly located. Initially, font samples of all expected characters are required and used for identification. Further studies suggest the use of character primitives to eliminate the need for a large font database. Heuristic and optimal primitives are considered and tested against a specific component class.

In this evaluation contrast-detail diagrams were used to measure the effects on observer performance of several image processing algorithms such as edge enhancement, smoothing and histogram equalization. The observer task was the detection of disk images in a uniform background. A new digital system, DigiRad System One, was used to digitize and postprocess images of a "Rose-Burger" contrast phantom. Results show that detectability of low contrast objects may increase for images taken with standard film/screen combinations. Anticipated changes on the System One may enhance its capabilities significantly.

A new image enhancement technique has been applied to enhance LANDSAT-4 Thematic Mapper (TM) thermal infrared (IR) imagery, and has improved its effective ground resolution from 120 meters to 30 meters. The technique is based on an adaptive multi-band least mean squares (LMS) method for computing an optimal image estimate from reference images, and a frequency domain replacement step. The approach relies on the assumption that at a sufficiently small resolution, registered TM imagery data is locally correlated across the bands. Using the local correlation property, visible and IR reference bands are used to predict the thermal IR image data. The prediction estimate is then used to augment the spatial high frequency information in the original thermal data. Preliminary experiments enhancing 120 meter IR imagery (simulated degradation) to 30 meter resolution show that the rms error can be reduced by 80% in high edge detail regions.

An evaluation was performed to characterize a new digital imaging system, the System One, DigiRad, Inc. and its image processing capabilities. This evaluation included reproducibility, dynamic range, linearity and the effects on several image recording system. Results show that the System One dynamic range is limited to 1300 levels in the range of 0.0 to 2.8 optical density units. It was observed that within this range, the System One response was approximately linear and reproducible within -1%. The system spatial resolution limit is between 3.0 and 3.5 1p/mm. All processing algorithms applied to digitized images of bar pattern film images degraded the resolution as compared to the images with no post processing. Improvements in System One are expected and will be evaluated as soon as they are available.

We obtained a sequential series of time-resolved tomographic two-dimensional images of a 4.5-MeV, 6-kA, 30-ns electron beam. Three linear fiber-optic arrays of 30 or 60 fibers each were positioned around the beam axis at 00, 61°, and 117°. The beam interacting with nitrogen at 20 Torr emitted light that was focused onto the fiber arrays and transmitted to a streak camera where the data were recorded on film. The film was digitized, and two-dimensional images were reconstructed using the maximum-entropy tomographic technique. These images were then combined to produce an ultra-high-speed movie of the electron-beam pulse.

A 64 KBPS codec for digital transmission of full motion monochrome video, which makes allowances for synchronization (frame, field and line) is proposed. Various combinations of down-up conversions, temporal filters, motion compensated frame interpolation, transform/predictive coding, quantizer and geometrical zonal sampling are applied to video sequences with an objective to obtain optimum combination. Simulation results based on these techniques are reported. A viable scheme for teleconferencing at 64 KBPS is proposed.

This paper discusses the status of the activities of the International Consultative Committee on Telegraph and Telephone (CCITT) in preparing technical recommendations for international video teleconferencing systems. It presents a number of driving factors which influence the preparation of recommendations, and discusses the salient technical aspects of the work. The efforts to date, which have resulted in preparation of three new draft recommendations, have concentrated on the development of a hypothetical reference connection for long distance interconnections, selection of interconnect bit rates, and in the selection of the types of video codecs to be employed. A hypothetical reference connection (HRC) which shows interconnection of a 2.0 Mb/s region with a 1.5 Mb/s region indicates that the transmission bit rate used on the international portion would be 1.5 Mb/s. The CCITT Study Group is also considering interconnect bit rates at multiples and submultiples of 384 kb/s. The Study Group has selected interframe type codecs for some applications and is currently studying both interframe and intraframe codecs for use in teleconferencing systems.

Over the past ten years many mathematical models of the human visual system (HVS) have been proposed for image processing applications.1-8 It has become clear that incorporating factors accounting for human perception can significantly improve overall picture quality. The purpose of this paper is to address the importance of applying the HVS model to image processing. In order to obtain good image quality at low bandwidth, it is proposed that a physiologically-based HVS model be incorporated with image compression systems. The physiological model needs no adjustment for input image as is required in the psychophysical model.

An efficient algorithm is proposed that computes the coefficients of a higher-order discrete cosine transform (DOT) from the coefficients of a lower-order DCT. The main feature of this algorithm is that it calculates the DCT coefficients of the larger block sizes from the DCT coefficients of smaller block sizes without any need to generate the inverse DCT. It is more efficient than the standard approach, which involves an inverse as well as DCT, if the standard fast-Fourier transform (FFT) procedure is used. However, it is not as efficient as the standard procedure if DCT and IDCT are generated using the newly discovered Winograd Fourier transform algorithm. This paper develops the computational algorithm for the one-dimensional DCT, then extends it to the two-dimensional DCT.

This paper describes a video compression technique which utilizes the alternating projection theorem for convex sets. The image to be transmitted is determined to be in certain convex sets and parameters defining these sets are sent. The receiver can then use the method of successive projections to locate an image which is in the intersection of the sets. If the intersection is small then the image determined should be close to the desired image. The coder can be made more robust by easily adding additional convex sets or using it in conjunction with other coding schemes such as motion compensation.

The rapid advance in digital electronics during recent years has enabled the real-time hardware implementation of many basic image processing techniques and these methods are finding increasing use in both commercial and military applications where a superiority to existing systems can be demonstrated. The potential superiority of an entirely passive, automatic image processing based navigation system over the less accurate and active navigation systems based on radar, for example "TERCOM", is evident. By placing a sensor on board an aircraft or missile together with the appropriate processing power and enough memory to store a reference image or a map of the planned route, large scale features extracted from the scene available to the sensor can be compared with the same feature stored in memory. The difference between the aircraft's actual position and its desired position can then be evaluated and the appropriate navigational correction undertaken. This paper summaries work carried out at British Aerospace Hatfield to investigate various classes of algorithms and solutions which would render a robust image matching system viable for such an automatic system flying at low level with a thermal I.R. sensor.

A hardware system for real-time image processing is described. The hardware is designed to work at a 10 MHz pixel rate, and will accommodate 512 x 512 images at 25 frames per second. The techniques of parallelism and pipelining are used to achieve the required speed. The system's primary function is the application of an edge detection algorithm to video data in real-time, the algorithm being a 2 x 2 Roberts operator thresholded with a function of local average brightness. The output of the system is a single bit/pixel edge picture which can be directly transferred into the memory of the host computer for further processing. Manipulation of the threshold function allows the system to disregard the edge detection function, and work directly on a greyscale image. The resulting single bit/pixel images can range from simple thresholded greyscale to the display of greyscale intensity contours. The edge detection process, coupled with these additional operational modes, makes the system a powerful tool for image processing.