Extensive research has shown that including target aspect angle measurements from an optical sensor can significantly improve the performance of radar tracking systems. Integrating sequences of target imagery with the kinematic information involves sets of image processing and sensor data fusion algorithms. A workstation has been developed to expedite the analysis of the algorithms and to integrate the image processing with selectable extended-state tracker modules. This workstation can access analog video imagery from a video optical disk controlled by a PC, segment the target in the image, and perform target identification and aspect angle estimation using a database of target models which span the range of possible aspects. The angle information is then `fused' with kinematic data to augment the tracker state estimator. The workstation is implemented with a powerful visual user interface in a UNIX/X- Windows environment, and includes a wide array of image and signal processing algorithms. Interactive modifications of processing sequences and `what if' analyses are easily conducted. The workstation provides a consistent user interface across a variety of applications. This system has also been used to implement phase retrieval and related image recovery algorithms.

This paper describes an environment that builds on UNIX for image processing on Sun workstations. It includes discussions on writing image analysis functions using a command line interface and on the manipulation of images from within an imaging program. It then describes the program imagetool, which displays images in a window environment and supplies a menu interface to the image-analysis functions, driven by the command line.

Traditionally, the quantity of wood in a truckload of logs is evaluated by a manual measurement of the enclosing volume. This paper describes the method developed at LNETI for measuring automatically the volume of a consignment of logs. The PC-based central unit processes the images from several solid-state medium-resolution cameras, using the software developed. The computer vision system, the image processing algorithms implemented therein, and the results obtained are presented.

Machine vision applications require fast, robust object descriptors. Area Parameter Acceleration offers these, based on seed parameters and connectivity from hardware. This strong approach to object analysis and machine vision will get stronger with upcoming Area Parameter hardware products.

Objects to be handled in medical image processing include large value arrays (curves, images, ...) and tightly associated data (patient's identification, conditions of acquisition, ...). Their validation constraints and structure are complex because they are dependent on the source and processing. The number of objects may change as new generators (source or processing) emerge. These objects and the constraints are formally described (BADAOUI9O), according to the following concepts: the basic type (for atomic attributes), the constructed type (for the multi-attribute features), the class type (for the complex entities and constraints) leading to a non-normalized relational model. This model, implemented on top of an RDBMS (Relational Data Base Management System) by way of a dictionary (metadatabase), defines entirely the base by intention and monitors its evolution. A tool associated with this model sets up the bridge with the relational model.

The remote detection of minefields has a high priority with the armed forces of several countries, including Canada. The Defence Research Establishment Suffield and the University of British Columbia are jointly developing methods to identify surface-laid minefields by real- time analysis of airborne sensor imagery. Passive multispectral imaging has been investigated, but the problem of analyzing monochromatic images from active infrared scanners is considered here. Among the severe constraints are that one must find clusters of bright regions each only several pixels wide, sparsely distributed over large areas. Also, the sharply forward peaked scattering from the mines and variable terrain roughness leads to the result that some mines are not detected. The required very high input data rate suggests simple, high speed target cuing which is amenable to hardware implementation. This is incompatible with slower image understanding techniques needed to identify mines by their shapes, and minefields by the spatial relationships between mines. A hierarchical algorithm is thus presented as a solution to the problem. In progressing from lower to higher levels, computational operations become more complex and less able to be implemented on conventional high-speed devices but the volume of data to be analyzed decreases. At the lowest level, non-suspect regions are rejected to drastically reduce the data rate. At the next higher level, suspect regions are segmented into homogeneous sub-regions. The next level consists of extraction of morphological features of the sub-regions. One level higher, sub-regions are classified based on extracted features. At the top level, spatial relationships between `mine-like' regions are determined and are used by knowledge-based methods to classify the imaged area as being a minefield with a specified likelihood. The algorithm, which includes a novel region segmenter, has been developed up to and including sub-region classification. Studies using simulated yet realistic thermal infrared minefield images have demonstrated that the algorithm is successful in detecting individual mines. Current efforts to make the algorithm execute in real-time, by implementing it on a transputer network communicating with array processors, will also be discussed.

The affine invariant recognition of two-dimensional objects which have the same shape but a different gray level function is a difficult problem in computer vision. A possible solution to the problem is based on the method of affine invariant moments. However, the moments have a badly conditioned dynamic range and they turn out to be very sensitive to noise. This paper presents a new method which is based on affine invariant Fourier descriptors for planar curves. It suggests a weighted parameterization for the boundary of an object based on its gray level function. The parameter of a boundary point is defined as the sum of gray levels of the area covered by a line from the center of mass to a starting point of the boundary moving along the boundary. The parameterization is an affine mapping with respect to the transformed object and to an arbitrary starting point of the contour finder. This parameterization enables the development of affine invariant Fourier descriptors. The Fourier descriptors consider two features, the contour and the gray level function of the object. The set of Fourier descriptors which characterizes the image pattern is not complete but their capabilities for pattern separation is very good. The method is less vulnerable to noise than the method of affine invariant moments. Application tests under real conditions demonstrate the practicability of the proposed algorithm. The complexity of the algorithm for a pattern with N pixels is 0(N).

We present a regularized pseudoinverse method which has successfully superresolved passive millimeter-wave images and that can be performed in the frame time of the imager.

Vision-based sensors detect a variety of defects with comparative ease, and inspect 100% of the material. Current commercial capabilities using Area Parameter feature extraction techniques are examined. Hardware used in these systems is reviewed with improvements in feature extraction bandwidth considered.

There has been a desire for some time by practitioners and researchers in the ophthalmology field for an instrument to automatically measure eye alignment quickly and accurately. With MS DOS based Personal Computer image processing systems becoming readily available and inexpensive, a method is presented that uses image processing tools to easily diagnose eye misalignment and initiate the appropriate treatment. The MS DOS based Personal Computer image processing system discussed is able to accurately measure the monocular inner pupillary distance, compute the axial length of each eye, and measure angular deviations so that tropias and phorias can easily be measured. Calibration of the imaging system requires that the pixel displacement as a function of eye movement be known. It is the purpose of this paper to present experimental data collected over several patients with and without strabismus (ocular deviations) and a detailed theoretical analysis that calibrates the imaging system discussed. In designing the imaging system to measure ocular deviations, several calibration schemes were developed. The first uses estimated data collected and published in the literature on the axial length of the eye as a function of the patient's eye (Estimated), the second uses the axial length of the eye measured from an A-scan (Calibrated), and the third uses the imaging system directly to compute the axial length (Physiologic).

A recently developed class of digital filters known as morphological pseudoconvolutions are applied to scanning tunneling microscopy (STM) images. These filters use morphological filtering to improve the characteristics of both moving mean and moving median filters. They filter equally in both the x and y directions, so as not to introduce artifacts, and they have an adjustable parameter that allows the user to restore the observed image completely as the parameter tends to infinity. Very few assumptions are made concerning image and noise content, only the shape of typical data being taken into account. These filters are shown to outperform, both visually and in the mean square error (MSE) sense, previously introduced Wiener filtering techniques. The filters are compared on typical STM type images, using both modeled and actual data. The technique is general, and has been shown to perform very well on all types of STM and Atomic Force Microscopy (AFM) images.

Algorithms for visual inspection of LSI wafer multilayer patterns have been developed. These algorithms compare corresponding images of two dies on a wafer. In this paper, two algorithms are proposed. The derivative-polarity comparison algorithm compares the polarities of the first derivatives of two images, and recognizes the regions whose polarities are not matched as positional discrepancies (defects), in order to cope with gray-scale differences caused by pattern thickness errors. The multiple-displacement pattern matching algorithm executes the above polarity comparisons at several positions with images suitably aligned, and determines the common unmatched regions as defective, in order to handle the interlayer- registration errors encountered with multilayer patterns. These algorithms were evaluated experimentally, and it was verified that defects of 0.3 micrometers or more can be reliably detected in multilayer patterns by combining these algorithms.

We present an algorithm for the joint estimation and segmentation of optical flow from gradient data, based on considerations from image grey level segmentation. Some first simulation results are shown and compared to existing spatial domain flow estimation algorithms. Promising results as well as strong indications for future enhancements are obtained.

Currently there is rapid progress in the development of model-based object recognition techniques. Current techniques and capabilities are reviewed and an example of model-based recognition for aerial reconnaissance is described.

For several useful tasks in photogrammetry and in model-based vision, this paper develops noniterative methods that require only the inversion of systems of linear equations. The methods are based on the theory of projective invariants. The following tasks are addressed: (a) Resection, or determination of parameters of acquisition geometry (requires six control points); (b) Intersection, or determination of the position of an object point from several images; and (c) Transfer, or model matching, which uses the image coordinates of a ground point in two images to predict the coordinates of that point in a third image in the presence of several other tie points in the three images.

Multiband synthetic aperture radar (SAR) is beginning to join the suite of spectral data sets available for feature discrimination and classification. For myriad practical applications, it is useful to quantify how well multiband SAR performs on its own, and how well it performs compared to and in conjunction with visual and infrared (IR) wavelength band sets. This paper examines these issues in the context of land use determination for a rural area in California's Central Valley, using multifrequency, multipolarization imagery from the NASA/JPL Airborne Imaging Radar (AIR). A supervised classification of the major terrain types is used to assess the performance of the AIR bands. Also, the best band combinations for feature discrimination are selected from a combined band set containing AIR, visual wavelength, and IR bands.

Considerable resources are devoted to the acquisition, storage, and measurement of stereo imagery for the purpose of extracting terrain elevation data. Improvement in the accuracy of stereo parallax measurement allows a reduction in the resolution of images collected and measured to obtain a required elevation accuracy. A reduction in resolution requirements can lead to significant savings in collection, storage, and processing costs. A neural network is described which uses image cross correlation data to determine stereo parallax disparity to fractional pixel accuracy at each pixel location in the images. Correlation values are obtained between image windows in the left view with a succession of overlay window positions in the right view. High resolution parallax disparity determination requires small image windows. Correlation peak locations for the small windows are often unreliable match points due to noise and relative parallax distortion between the images in regions of rapid elevation variation. Further processing of the correlation data is needed to reduce errors. A neighborhood of correlation data is input to a neural network. The network outputs the parallax disparity for the pixel location (in the left view) centered on the neighborhood. The network was trained and tested using stereo images of a scene containing a variety of hills, terraces, flat terrain, and roads. The parallax disparities for this stereo pair were carefully measured manually at high resolution. These measurements provided accurate sub-pixel accuracy at the lower resolution used to train and test the network. A network with two `hidden layers' was trained using the backward-error-propagation method. Network results on separate test sections of the scene show significant improvement over results using alternate methods.

A new generation of remote sensing instruments, called imaging spectrometers, are designed to collect image data in literally hundreds of spectral channels simultaneously, providing significantly enhanced amounts of information as compared to existing systems for studying biophysical processes. The advantage of such high-dimensional data comes at the cost of increased data complexity. The volume and complexity of the new data, in turn, presents a challenge to the traditional image analysis methods and requires that new approaches be developed to allow rapid and effective analysis of the imagery. This paper describes a technique which reduces the data dimensionality, while retaining sufficient pertinent information that the original high-dimensional signals provide for class separation. Results of applying this technique to 224-band AVIRIS data are presented.

This paper describes a method for learning the 2-D shapes of the objects from their sample image clips. We view an object as a congregation of a set of component parts with simple shapes. When presented with a sample image clip of an object, our learning system first detects the components of that object and saves the shape descriptions of those components in the object model. Next it determines the geometrical relationships between the components which are also saved in the object model as constraints. Finally, it generates a strategy for recognizing such an object. As we will show, our system can use this automatically extracted information to detect such an object in other scenes, even when the object is partially occluded.

Computer pattern recognition has been expanding its application in various fields. Its application in printed circuit board (PCB) inspection, for example, will accordingly reduce the requirement of manpower as well as the error rate, if equipped with adequate accessories; real time processing can then be implemented with defect locating feature. The design philosophy of this system is focused on how to establish a complete PCB inspection methodology with an error indication for a more effective amending process. The algorithm of this system is as follows: first, use the CCD camera to scan the PCB under inspection, then compare it with the correct one, which has already been stored in the computer memory, and compare the results. If any error exists, then a `+' will be marked on the defect spot from the computer, and overlapping will be executed via a computer-controlled zooming LCD (liquid crystal display) overhead projector. The `+' mark can specify the defect spot clearly. The size and the position of the overlapping image with the tested PCB will be exactly one to one. The optical distortion error is corrected by computer program using the moire and optical analysis method.

We study machine vision post-inspection of products for automatic correction of manufacturing processes. The general question is posed as a `discrete-event' stochastic control problem with the camera as the feedback sensor. The physical models employed yield insight into design procedures for robustness.

This study explores the application of digital image processing techniques to a machine vision system for log inspection in the forest products industry. This machine vision system uses the computerized tomography (CT) imaging to locate and identify internal defects in hardwood logs. To apply CT to these industrial vision problems requires efficient and robust image processing methods. Several image processing techniques are addressed in this paper: adaptive image smoothing, multi-threshold-based segmentation, morphological filtering, and 3-D connectiveness labeling. Experimental results of these image processing techniques with CT images from two different wood species demonstrate the efficacy of the inspection system.

This paper describes the image processing methodology developed in order to implement a machine for automatically punching corks in an optimized manner from raw cork strip. Optimization of the use of cork is important since the manufacturers are paid for the number and quality of the corks they produce. Currently the punch-operator decides where to punch corks based on a quick visual inspection of the strip. The machine being developed uses both CCD cameras and a laser scanning system to acquire information about the distribution and nature of defects in the strip. With this information the system is able to decide on a punching pattern which optimizes the value of the corks produced by the strip. The paper presents in some detail the image processing and punching-pattern optimization algorithms used (which are based on morphological methods) and also describes the PC-based hardware employed during their development.

A system is under development in which surface quality of a growing bulk mercuric iodide crystal is monitored by video camera at regular intervals for early detection of growth irregularities. Mercuric iodide single crystals are employed in radiation detectors. A microcomputer system is used for image capture and processing. The digitized image is divided into multiple overlapping sub-images and features are extracted from each sub-image based on statistical measures of the gray tone distribution, according to the method of Haralick. Twenty parameters are derived from each sub-image and presented to a probabilistic neural network (PNN) for classification. This number of parameters was found to be optimal for the system. The PNN is a hierarchical, feed-forward network that can be rapidly reconfigured as additional training data become available. Training data is gathered by reviewing digital images of many crystals during their growth cycle and compiling two sets of images, those with and without irregularities.

A new method in human motion analysis is presented for overcoming the problem of the shifting of skin-mounted position markers relative to the skeleton. The present version of the method is based on two-dimensional video processing and involves the recording of subjects wearing special clothing. The clothing is designed in such a way as to permit the unambiguous spatial shape recognition of each of the 17 body segments by means of an edge detection algorithm. The latter and the algorithms for the computation of segment translation and rotation constitute improved versions of previously used algorithms, especially with respect to the execution times of the respective computer program on ordinary PCs. From the recognized shapes, the translation and rotation of each segment relative to its initial configuration is computed by using positional information from the previous frames. For the first frame to be analyzed, a starting algorithm has to be applied. Finally, the configurational coordinates of the body model are calculated from the respective spatial linear and angular positions.

The inspection of surface-mount technology printed circuit boards is an increasing problem due to the fine pitch technology that is beginning to be used. Human visual inspection is unsuitable for this technology necessitating the use of computer-based inspection systems. Correlation has been used by the authors to detect solder bridges and misaligned leads on surface mount components. This operation is too slow on a general purpose computer requiring over seven seconds to inspect each component lead. A digital signal processor based multiprocessor system has been designed to provide the required computational power. Using optimization techniques and restructuring the data to be correlated enables each processor in the system to inspect 424 leads per second.

It is well known that the Hubble Space Telescope suffers from spherical aberration. Much that is scientifically valuable can be done with the telescope in its present condition, but we must install corrective optics. An analysis of images gives the same results as measurements on the equipment used to fabricate and test the primary mirror, so such optics can be designed with confidence, with a derived conic constant on the primary mirror of -1.0139(5). However, the optics assembly and integration with the spacecraft poses great challenges because if the corrected images are to be diffraction limited, they must be aligned to better than one percent of the beam diameter. Some other residual effects of the spherical aberration will remain after installation of the corrective optics, primarily in the pointing and collimation of the telescope. We summarize the present imaging performance of the observatory, and compare it with the expected performance when corrective optics (COSTAR and WFPC 2) are installed on-orbit.

After the Hubble Space Telescope was stabilized so that it could image a star, scientists were dismayed to find that the telescope produced severely degraded images. A flaw in the manufacture of the primary mirror causes an aberration that appears to be circularly symmetric. In this paper we review some of the algorithms and procedures that have been used to perform image restoration of the Hubble data.

The spherical aberration of the primary mirror of the Hubble Space Telescope has seriously reduced the sensitivity and degraded the resolution of images and spectra from the instruments onboard the satellite. In this paper, we describe the nature of the aberration problem and the resulting point-spread functions for both the imaging and spectrographic detectors. A number of image restoration algorithms have been evaluated for use on HST data, including both linear and non-linear techniques, and sample results are shown for a modified Richardson- Lucy method. The best prospects for obtaining improvements in the restored images are likely to derive from developing techniques for adaptive regularization.

It was found, on orbit, that the Hubble Space Telescope had a conic constant error in the primary mirror. The result of the error is a substantial amount of spherical aberration in the image, significantly reducing image resolution and encircled energy. Parametric phase retrieval was the method used to determine the source of error and to find the magnitude from on-board camera images. The parameters which are estimated are a set of annular Zernike polynomials which are analogous to the classical aberrations. This was one of the first practical uses of phase retrieval for on-orbit measurement. This paper contains an overview of the algorithms, how they were used and the major results.

Phase retrieval algorithms, including the iterative transform algorithm and gradient search algorithms, were generalized to include the effects of propagation through a complicated optical system and to discount the effects of bad CCD pixels. For the gradient search algorithms, analytic gradients were derived that greatly speed up the computation over finite difference methods. For the Hubble Space Telescope (HST), the aperture function was reconstructed and the phase errors were retrieved. This information is useful to design correction optics for the telescope and for the deconvolution of blurred images from the HST.

The first objective of this study is to analyze the use of neural networks for bandwidth compression. This is achieved by developing a neural network algorithm using the Kohonen self-organization technique to perform vector quantization. The second objective of the study is to combine the neural network vector quantizer with a DPCM encoder for more efficient bandwidth compression. The bandwidth compression techniques are simulated and their performance is evaluated using one-dimensional wideband signals. Vector quantization (VQ) has proved to be an efficient method for bandwidth compression of both one-dimensional signals and imagery data. The reason VQ is not utilized in many practical applications is due to the fact that the performance of VQ becomes superior to other techniques such as transform coding and DPCM only for large vector dimensions. Large vector dimensions also increase the number of computations (per sample) and the memory requirements of the VQ; the increase is an exponential function of the vector dimension. For this reason, in this study VQ has been used in combination with other methods of bandwidth compression so that using a small vector dimension still improves the overall system performance. Neural networks present a parallel approach to data classification that may simplify the architecture of the classifier in the VQ, thus making vector quantizers with large dimensions more practical. In this study, we have developed a neural network algorithm for vector quantization which is based on Kohonen self-organization technique. In the following we discuss the neural network classifiers and their utilization in vector quantization of the DPCM encoders. Simulation results showing the system performance of these systems for one-dimensional modulated signals is presented and the results are discussed.

A numerical technique is described for designing multiresolution, maximally decimated, parallel-structured, one-dimensional filter banks. The filter banks have properties which make them well suited to image subband coding applications. In particular, they provide nearly perfect reconstruction, are nearly orthogonal, provide joint localization in the spatial and spatial-frequency domains, and allow no leakage of dc into higher subbands. This paper describes the technique and presents a design example.

We consider the compression of multispectral images using Hilbert scanning and adaptive arithmetic coding. The Hilbert scan is a general technique for continuous scanning of multidimensional data. Arithmetic coding has established itself as the superior method for lossless compression. The aim of this paper is to investigate the integration of the arithmetic coding methodology and a n-dimensional Hilbert scanning algorithm developed by Perez, Kamata and Kawaguchi.

The problem of easily implementable filter bank design for subband coding is discussed. A technique is proposed to generate filter banks with coefficients in powers-of-two and with maximum localization. The hierarchical structure of the filter bank is then used to eliminate the temporal redundancies by using a multiresolution motion estimation technique. Simulations show a robust estimation of the motion field with a relatively simple implementation.

The block-based motion estimation algorithm we have previously proposed possesses the advantages of a high estimation accuracy and a good prediction performance. It can efficiently be employed in a motion-compensated DPCM coder to minimize the data bits for the transmission of prediction errors. However, the motion overhead data rate, although it has been compressed dramatically by the coding method we proposed, is still too high to work at very low bit-rates like 64 kb/s. In this paper, a hierarchical motion-compensated interframe DPCM algorithm is presented to cope with low bit rates coding. In the meanwhile, the motion- compensated prediction error images are encoded by a hierarchical quadtree structure coder, which is more efficient and simpler than the commonly adopted DCT transform coding for the prediction errors. Experimental results show that an excellent performance has been obtained for very low bit rate coding. The signal-to-noise ratios of the reconstructed luminance images on average are about 37.5 to 38 db at 0.05 bits/pel and 38.7 to 39.5 db at 0.09 bits/pel for the sequences examined.

Shape analysis and synthesis are important capabilities in many image processing applications such as scene analysis, computer-aided design, and cartoon generation. A very important aspect of shape analysis is proper representation of object boundaries. Chain coding is an efficient, often-used method of representing these boundaries. The conventional method to generate a chain code is: (1) find and select an object boundary pixel; (2) find the nearest edge pixel, code its orientation, and mark it as used once; (3) repeat step (2) until there are no more boundary pixels. This procedure is computationally expensive. The bottleneck of the process is determining the direction to the nearest edge pixel. In the worst case, six of the eight surrounding pixels must be checked for 8-neighbor connectivity. We present a new, fast method for determining the connectivity for each pixel in the entire image using a 3 X 3 convolution kernel that produces an oriented connectivity map for the entire image. The chain code is then generated by following the map. The significant advantage is the ability to exploit high-speed convolutional processors such as HNC's Vision Processor (ViP). Using the ViP, the necessary convolution can be accomplished in less than 7 milliseconds for a 512 X 512 image. The ViP also can perform most other image processing functions within 7 milliseconds. Here we introduce the chain code algorithm based on a convolution result.

An improved method for the reversible compression of industrial radiographs is described. The method uses one of the multiresolution decorrelation techniques, viz., hierarchical interpolation (HINT), to decorrelate a given image. However, instead of encoding the decorrelated image pixels at each resolution level using a memoryless source as in the traditional HINT method, a statistical source model with multiple contexts is employed and the pixels are encoded using the appropriate contextual statistics. Experiments on industrial x-ray images show that the improved method achieves about 19% more compression on the average than the traditional method.

Two new filters for image enhancement are developed, extending the early work of the authors. One filter uses a new nonlinear time dependent partial differential equation and its discretization, the second uses a discretization which constrains the backwards heat equation and keeps it variation bounded. The evolution of the initial image as t increases through U(x,y,t) is the filtering process. The processed image is piecewise smooth, nonoscillatory and apparently an accurate reconstruction. The algorithms are fast and easy to program.

The quantification of the textural differences between regions and the classification of subregions by their texture is an important component of image analysis. Here we demonstrate that texture quantifiers based on the wavelet transform of an image are good discriminators of texture differences.

Here, we describe a new and unique image sharpening method based on computational techniques developed for CFD. Our preliminary experience with this method shows its capability for nonlinear enhancement of image edges as well as deconvolution of an image with random noise. This indicates a potential application for image deconvolution from sparse and noisy data resulting from measurements of backscattered laser-speckle intensity.

Advanced filter algorithms using in-plane rotation invariance as an additional constraint during filter synthesis are presented. The new rotation-invariant Minimum Average Correlation Energy (RI-MACE) filter provides sharp, easily detected correlation peaks and excellent discrimination against other false objects. In the presence of background noise and clutter, the rotation-invariant Minimum Noise and Correlation Energy (RI-MINACE) filter (modified RI- MACE filter) uses the noise information during filter synthesis to obtain improved noise performance and also maintain easily detected correlation peaks. New test results are presented to show the improved discrimination capability of the RI-MACE filter and the improved noise performance of the RI-MINACE filter.

A new approach to intensity-invariant filter design is introduced. This approach post-processes a distortion-invariant correlation response. The correlation responses of a filter bank are mapped into a 3-D volume matrix where the third dimension is intensity. For each given intensity level the volume matrices are weighted by a complex exponential with unit magnitude and linear in phase. These weighted volume matrices are summed together to form filter volume matrix. Pages or slices from the volume are used as a bank of 2-D correlation filters. Computer simulations are included to quantify the performance of this approach.

This paper presents a novel approach for estimating the 3-D translation of a rigid surface using a binocular observer. The method uses the general theory of moments and is based in the aggregate stereo approach. The computation of a limited set of moments of four images is necessary. Real-time implementation of the algorithm is feasible since image moments can be computed at frame rates very easily using parallel architectures.

Detecting targets from infrared (IR) images in the absence of a priori information is a very difficult task. In this paper, we present an unsupervised detection algorithm based on Gabor functions for detecting targets from a single IR image frame. The only explicit assumption made is that the targets can be considered a rectangle. The algorithm consists of three steps. First, it locates potential targets based on a rectangle matching pattern by using low resolution Gabor functions which resist noise and background clutter effects. Then it removes false targets and eliminates redundant target points based on a similarity measure. These two steps mimic human vision processing but are different from Zeevi's Foveating Vision System. Finally, it uses both low and high resolution Gabor functions to verify target existence. This algorithm has been successfully tested on several IR images that contain multiple examples of military vehicles and aircraft with different size and brightness in various background scenes and orientations.

In this paper we address the problem of removing blur from, or sharpening, astronomical star field intensity images. A new image restoration algorithm is introduced which recovers image detail using a constrained optimization theoretic approach. Ideal star images may be modeled as a few point sources in a uniform background. It is therefore argued that a direct measure of image sparseness is the appropriate optimization criterion for deconvolving the image blurring function. A sparseness criterion based on the lp quasinorm is presented and algorithms for sparse reconstruction are described. Synthetic and actual star image reconstruction examples are presented which demonstrate the algorithm's superior performance as compared with the CLEAN algorithm, a standard star field deconvolution method.

This paper describes improved methods for enhancing low resolution images with the aim of extracting the desired image from a noisy background, while preserving its surface features. A novel modified Laplacian filter (Laplace-8) was used. This was based on the classical Laplacian filter (Laplace-4) and operates in a similar fashion by using the Laplace coefficient to determine the level of enhancement. Laplace-4 has a major shortfall in that it only enhances gradients, and so the gray levels of a smooth image surface are left unchanged. This results in severe image surface distortion. The improved filter (Laplace-8), aims to alleviate the distortion by enhancing the whole surface of the image as well as producing good contour enhancement. This is possible even if the image surface is totally smooth. Two controlling techniques were used to compare the two filters, namely, limited gray level (limited between 0 -255) and unlimited gray level. Results were based on the correlations of original and Laplace filtered images, and the statistics of their contours and surfaces. The images produced by Laplace-8 filtering were shown to be superior to that of Laplace-4, showing little image distortion in the case of unlimited gray level enhancement. Laplace-8 is also an effective contour extractor, producing higher contour gray levels for a given enhancement (Laplace) coefficient. The paper describes in detail the performance of the Laplace-8 method with aid of examples.

In Japan, HDTV broadcasting will start this year. Point-to-point HDTV live demonstrations have also been carried out on an experimental and commercial basis for several years. To realize these services, bit rate reduction coding is one of the key technologies. This article describes the present state of development of HDTV digital codecs in Japan.

The studies of HDTV systems in Europe have been driven by the concern of a compatible evolution from conventional definition TV to HDTV. This approach has led to the D2- HDMAC packet broadcasting system which is now under development and compatible with the D2-MAC packet system. For professional (contribution) applications digital codecs able to transmit HDTV over 140 Mbit/s links have also been developed and a research program was launched several years ago to address the specific issue of TV and HDTV distribution over digital optical fiber networks such as the future B-ISDN. Again, a compatible approach has been favored for this application with the goal of developing a multiresolution coding system enabling simultaneous decoding of signals of various resolutions. In this paper, the results achieved so far in designing the relevant systems are presented.

Digital techniques are widely adopted to process the video signal, and nowadays there is the necessity of transmitting conventional and high-definition television signals between different studios. In the near future it could be possible to deliver a digital signal to the consumer. To allow the use of the presently available digital networks and satellites, sophisticated compression techniques have been devised to limit the bit-rate requirements and to provide a high-quality and reliable service. A hybrid predictive/transform system has been devised and implemented in the framework of the European project EU 256. The main parameters of this system are in accordance with those being recommended by ETSI and CMTT for the transmission of conventional component TV. Codecs are available for the TV and HDTV formats presently in use and can operate with a wide range of transmission rates. The optimization of the system and the evaluation of its performance have been carried out on the basis of a large number of subjective tests in accordance with the user requirements specified by the standardization bodies. The codecs have been extensively tested on-field during experimental point-to-multipoint satellite transmission of HDTV signals on the occasion of soccer matches.

To realize the complex algorithms of encoding and decoding HDTV signals at the required pixel rate, it is advisable to split the image into several data channels. These channels have to be processed in parallel. As an alternative to splitting the image into horizontal or vertical stripes, the image can be split into different frequency bands. In this paper, a coding scheme for HDV signals, based on band splitting and band interpolation, is presented (the abbreviation HDV is used for high-definition video, symbolizing the use of the presented codec for more than television purposes). The concept of subband coding in combination with motion compensated prediction and transform coding is described.

The problem of Image Super-Resolution is defined as processing an image in such a way as to extend the bandwidth of the image beyond the diffraction-limited spatial frequency imposed by the optical aperture forming the image. We present an algorithm which is derived by MAP estimation under Poisson modelling assumptions. The algorithm has been demonstrated empirically as capable of producing Super-Resolution. We hypothesize that the modelling process is responsible for adding additional information that constitutes the potential for image bandwidth extension. We outline an information theory analysis to quantify the modelling information, and indicate the forthcoming efforts to be conducted in pursuit of this analysis.

This paper is concerned with the estimation of the image motion field from a pair of consecutive noisy frames. The maximum likelihood principle is invoked for estimating the nonrandom but unknown displacement function. In our developments, we consider processing both of the observed images (jointly) through a 2 X 2 noncausal matrix filter. The design of this matrix filter depends on the assumed values of the parameters for the displacement function. The analysis presented is the extension and generalization of the work originally established by Stuller who studied the problem of maximum likelihood estimation of variable time delay. The developments are specialized to the case for which the motion field is modeled by an affine transformation. Simulations are performed which indicate the validity of the estimator in the presence of noise. Results of the simulations are presented.

A preliminary investigation confirmed the possibility of assessing the translational and rotational information content of simple artificial images. The calculation is tedious, and for more realistic patterns it is essential to implement the method on a computer. This paper describes an algorithm developed for this purpose which confirms the results of the preliminary investigation. Use of the algorithm facilitates much more comprehensive analysis of the combined effect of continuous rotation and fine translation, and paves the way for analysis of more realistic patterns. Owing to the volume of calculation involved in these algorithms, extensive computing facilities were necessary. The major part of the work was carried out using an ICL 3900 series mainframe computer as well as other powerful workstations such as a RISC architecture MIPS machine.

This paper describes an approach of feature extraction from range and intensity data for three- dimensional model-based target recognition systems. The range image of the object under study is acquired using a methodology based on a passive stereo-vision model. A coarse to fine strategy is used to explore the features in the images, and cepstral analysis provides information about disparities between corresponding areas at each step. The surfaces of the 3- D target recovered from the passive stereo-vision model are represented by aspect graphs which are used in vector form as the input for a neural network classifier.

Noisy objects, partially occluded objects, and objects in random positions and orientations cause significant problems for current robotic vision systems. In the past, an association graph has formed the basis for many model based matching methods. However, the association graph has many false nodes due to local and noisy features. Objects having similar local structures create many false arcs in the association graph. The maximal clique recursive and tree search procedures for finding sets of structurally compatible matches have exponential time complexity, due to these false arcs and nodes. This represents a real problem as the number of objects appearing in the scene and the model set size increase. Our approach is similar to randomized string matching algorithms. Points on edges represent the model features. A fingerprint defines a subset of model features that uniquely identify the model. These fingerprints remove the ambiguities and inconsistencies present in the association graph and eliminate the problems of Turney's connected salient features. The vision system chooses the fingerprints at random, preventing a knowledgeable adversary from constructing examples that destroy the advantages of fingerprinting. Fingerprints consist of local model features called point vectors. We have developed a heuristic approach for extracting fingerprints from a set of model objects. A list of connected and unconnected scene edges represent the scene. A Hough transform type approach matches model fingerprints to scene features. Results are given for scenes containing varying amounts of occlusion.

Imaging applications often require image warping in order to obtain a desired spatial effect. In implementing a warping technique, means for transforming and resampling the source data must be chosen. A forward or reverse transformation mapping can be selected, though interpolation is more accurate using the reverse format. When applying a noninvertible transformation, forward mapping becomes necessary. A technique called bin mapping introduces an intermediate storage mechanism which allows for precise interpolation while applying a forward transformation.

This paper investigates a new accurate method of automatic searching center measurement which is suitable for all kinds of profiles. Concepts about planar relative motion are introduced. The paper focuses on the introduction of the exact solution of measured value, prototype value, and deviation value, to the establishment of the objective function for valuating the deviation of profile of a line, and to the description of the algorithms of searching field and iterative search for `deviation value -- new center -- deviation value ...' The flow chart of the computer program and the results of the experiment are given.

In this paper, we present an inverse filtering technique for synthesizing holographic correlation filters for recognizing targets in the presence of distortion of the input image. Numerical simulations and experiments are done.

We describe an optical 3-D sensing system by phase measuring profilemetry (PMP) for a 3-D bottom hole pattern. The principle, system construction, and experimental results are described.

In this paper, a new method of 3-D object recognition using optical techniques and a computer is presented. We perform 3-D object recognition using moire contour to obtain the object's 3- D coordinates, projecting drawings of the object in three coordinate planes to describe it and using a method of inquiring library of judgement to match objects. The recognition of a simple geometrical entity is simulated by computer and studied experimentally. The recognition of an object which is composed of a few simple geometrical entities is discussed.

In this paper, we represent a new method for adjusting color of pseudo-color encoding image, The method is based on the technique for phase image pseudo-color modulated by graring and use a reflecting ray path to produce diffraction of dual phase graring at object plane of 4f system so that the pseudo-color at output plane can be adjusted.

In the process of image acquisition, the object of interest may not be evenly illuminated. So an image with shading irregularities would be produced. This type of image is very difficult to analyze. Consequently, a lot of research work concentrates on this problem. In order to remove the light illumination problem, one of the methods is to filter the image. The dc notch filter is one of the spatial domain filters used for reducing the effect of uneven light illumination on the image. Although the dc notch filter is a spatial domain filter, it is still rather time consuming to apply, especially when it is implemented on a microcomputer. To overcome the speed problem, a parallel dc notch filter is proposed. Based on the separability of the algorithm dc of notch filter, image parallelism (parallel image processing model) is used. To improve the performance of the microcomputer, an INMOS IMS B008 Module Mother Board with four IMS T800-17 is installed in the microcomputer. In fact, the dc notch filter is implemented on the transputer network. This parallel dc notch filter creates a great improvement in the computation time of the filter in comparison with the sequential one. Furthermore, the speed-up is used to analyze the performance of the parallel algorithm. As a result, parallel implementation of the dc notch filter on a transputer network gives a real-time performance of this filter.

This paper presents a new method of identification and restoration of images with out-of-focus blurs. The basic idea of the method is as follows: first use the gradient image of the blur image being processed to estimate the initial value of blur radius; then use the least square restoration technique and the Fibonacci optimal search technique to determine the blur radius. The experimental results with artificially and true defocused images showed that the present method is successful.

Transformation of the image motion vector to a different domain exhibits impressive properties, such as resistance to noise and reduction in computational time and storage requirements. Using the steepest ascent algorithm in a transformed domain, we develop an iterative technique for frame-to-frame image motion estimation. The estimator seeks for the maximum likelihood (ML) estimate of the transformed-coefficients of the motion field. The scheme is implemented using discrete cosine, discrete sine, slant, Hadamard-Walsh, and Haar transforms. Since the motion field is generally slowly-varying, it is shown that by ignoring the higher number transformed-coefficients, substantial improvement in noise reduction is achieved. Simulation experiments are performed to indicate the validity of the analysis for real images in the presence of noise. Results of the simulations are presented.

Autofocus methods in SAR and self-survey techniques in SONAR have a common mathematical basis in that they both involve estimation and correction of phase errors introduced by sensor position uncertainties. Time delay estimation and correlation methods have been shown to be effective in solving the self-survey problem for towed SONAR arrays. Since it can be shown that platform motion errors introduce similar time-delay estimation problems in SAR imaging, the question arises as to whether such techniques could be effectively employed for autofocus of SAR imagery. With a simple mathematical model for motion errors in SAR, we show why such correlation/time-delay techniques are not nearly as effective as established SAR autofocus algorithms such as phase gradient autofocus or sub- aperture based methods. This analysis forms an important bridge between signal processing methodologies for SAR and SONAR.

The growing interest and importance of high quality imaging has several roots: Imaging and graphics, or more broadly multimedia, as the predominant means of man-machine interaction on computers, and the rapid maturing of advanced television technology. Because of their economic importance, proposed advanced television standards are being discussed and evaluated for rapid adoption. These advanced standards are based on well known image compression techniques, used for very low bit rate video communications as well. In this paper, we examine the expected improvement in image quality that advanced television and imaging techniques should bring about. We then examine and discuss the data compression techniques which are commonly used, to determine if they are capable of providing the achievable gain in quality, and to assess some of their limitations. We also discuss briefly the potential of these techniques for very high quality imaging and display applications, which extend beyond the range of existing and proposed television standards.

We study non-linear signal processing for detection of threshold signals in chaotic backgrounds. The method is based upon synthesis of a polynomial neural network (PNN) for prediction of chaotic dynamics. Using a solution of the Mackey-Glass equation as a chaotic background, we synthesize a non-linear PNN processor which produces a threshold signal to background improvement of over 20 db.

Among the various image data compression methods, the discrete cosine transform (DCT) has become the most popular in performing gray-scale image compression and decomposition. However, the computational burden in performing a DCT is heavy. For example, in a regular DCT, at least 11 multiplications are required for processing an 8 X 1 image block. The idea of the scaled-DCT is that more than half the multiplications in a regular DCT are unnecessary, because they can be formulated as scaling factors of the DCT coefficients, and these coefficients may be scaled back in the quantization process. A fast recursive algorithm for computing the scaled-DCT is presented in this paper. The formulations are derived based on practical considerations of applying the scaled-DCT algorithm to image data compression and decompression. These include the considerations of flexibility of processing different sizes of DCT blocks and the actual savings of the required number of arithmetic operations. Due to the recursive nature of this algorithm, a higher-order scaled-DCT can be obtained from two lower-order scaled DCTs. Thus, a scaled-DCT VLSI chip designed according to this algorithm may process different sizes of DCT under software control. To illustrate the unique properties of this recursive scaled-DCT algorithm, the one-dimensional formulations are presented with several examples exhibited in signal flow-graph forms.

In many image processing applications, labeling objects and computing their features for recognition are crucial steps for further analysis. In general these two steps are done separately. This paper proposes a new approach to label all objects and computer their features (such as moments, best fit ellipse, major and minor axis) in one pass. The basic idea of the algorithm is to detect interval overlaps among the line segments as the image is scanned from left to right, top to bottom. Ambiguity about an object's connectivity can also be resolved with the proposed algorithm. It is a fast algorithm and can be implemented on either serial or parallel processors.

Hybrid combinations of morphological, feature extraction and pattern recognition methods are used to analyze disordered structures. As a simple example of a disordered structure, we consider locations of a number of rectangles in an image and determination of the distributions of orientations and lengths of each. Applications of our techniques to textured patterns are then addressed. New elementary and directional structuring elements are used.

In a previous paper, the authors demonstrated how image warping could be implemented using a pair of two-dimensional (2-D) spatial look-up tables. In that paper, the focus was on generating the tables, not on antialiasing the results. In this paper, the authors present a method for generating resampling kernels from the spatial look-up tables for the purpose of reducing aliasing. The method generates resampling kernels based on the reduction in spatial frequency content that must take place in order to produce an antialiased warped result. Since the reduction in spatial frequency required for each output pixel can be unique, the method is adaptive.