A new computational technique for digital image bandwidth compression is presented. This technique is based upon the truncation of an observed image. The original image is predicted using a non-parametric approach. This predictor is a random grain allocation (RGA) algorithm, an algorithm that is similar to a Monte Carlo method. Using the RGA algorithm and with an appropriate block partitioning of the original image, significant image bit reduction can be achieved. Several image bandwidth compression examples are presented.

A K-th order adaptive transform coding algorithm is introduced to achieve a better tradeoff between the bi-trate and the image quality. This scheme first discriminates the image contents, based on the variations of the human visual degradation sensitivity on arious contents, and the appropriate parameters are then chosen accordingly. Some simulation results are given to illustrate the performance of this adaptive scheme. It is demonstrated that better image quality can be achieved with lower bitrate with this adaptive algorithm

Images that contain a high degree of complexity, such as natural landscapes, are difficult to compress well because of the large amount of information they contain. But, in some cases, this intricacy can be described by a simple set of rules. Such is the case with Fractal sets. Fractals with their properties of roughness and self-similarity, offer the best geometry for modeling certain highly-detailed images. A simple set of equations applied iteratively to themselves can generate a complicated digital image. Iterated Function Systems (IFS) [2] offers a method of describing complicated digital files with a small set of functions exhibiting fractal properties. The image to be coded with an IFS is first covered with affine transformations of itself. The coding is accomplished by saving the coefficients of the transformations. The decoding is performed by generating a dynamical system whose attractor is suitably close to the original image. The amount of distortion is dependent on the quality of the initial covering. This paper will describe the mathematics of IFS, the coding and decoding of a digital image with IFS, error analysis of IFS compression, and comparison to other compression techniques.

In this paper, a VLC decoder IC based on the table-lookup technique is described. In this design, a barrel shifter is employed to shift a number of bits corresponding to the decoded codeword. This parallel approach lowers the hardware speed constraint and reduces the system buffer memory size as well. The barrel shifter in conjunction with a codeword table and a word-length table makes the one-step decoding possible. The table storage required by this approach is the type of associative memory which is most efficiently implemented in Programmable Logic Arrays (PLAs). A Content Addressable Memory (CAM) approach that facilitates programmability is also addressed. Design considerations for multiple code tables and tables containing escape codes are discussed for more general applications of the decoder IC. The important issue of decoder interface circuitry, often ignored by others, is also fully considered. Such circuitry can greatly reduce the system complexity. Finally, the implementation of a mask programmable VLC decoder IC is demonstrated using a silicon compiler. Based on our simulations, the output symbol rate for a VLC decoder with 64 entries is about 30 MHz. This speed is high enough for most real-time applications including CCIR 601 digital video (Y: 13.5 MHz, U and V: 6.75 MHz each).

A new coding strategy for low bit-rate (64 Kb/s) image sequence transmission is presented. On the contrary to the conventional motion estimation methods in which only the immediate former frame is referenced for obtaining the motion patterns, more than one previous frames are used in the motion evaluation procedure to estimate the temporary motion behavior. Another sequence, which is called "the motion sequence", is estabished, alongside the image sequence. Through investigating and analysing the previous frame of the motion sequence the moving objects can be identified from the background and consequently two different image signals are separated. The coding operation is then controlled by the information from the motion sequence. The moving parts of the image data are coded in its normal frame sampling frequency while the background data can be treated coarsely by refreshing them in a much lower frequency. A finally reconstructed image frame is then an unchanged background superimposed by the decoded moving objects. This low bit-rate coding method is especially interesting in the applications such as teleconferencing and videophone in which the moving objects are usually the type of "head and shoulder" and the background data are usually unchanged.

The need for a compression system to significantly reduce the size of high resolution PC/AT based Video Graphics Array (VGA) image (picture) files, which can range from 64 Kbytes to over 1 Mbyte, stimulated an extensive research program that resulted in the development of new image conpression systems. This Discrete Cosine Transform (DC') based compression scheme can reduce VGA image file sizes by a factor of more than 6-to-1 with only minor detectable image degradation. The technology to be discussed is designed specifically to operate on the PC/AT-based platform equipped with VGA display. This paper will discuss the characteristics of VGA images versus full color RGB images and the compatibilities between these two types of images.

Efficient compression of image data requires the understanding of the noise characteristics of sensors as well as the redundancy expected in imagery. Herein, the techniques of Differential Pulse Code Modulation' (DPCM) are reviewed and modified for information preserving data compression. The modifications include: o Mappings from intensity to a user selected equal noise variance space, o Context dependent one and two dimensional predictors, o Rationale for non-linear DPCM encoding based upon an image quality model, o Context dependent variable length encoding of 2x2 data blocks, o Feed back control for constant output rate systems. Images have been processed at compression rates between 1.3 and 2.8 bits per pixel. The need for larger block sizes, 2D context dependent predictors, and the hope for sub-bits-per-pixel compression which maintains spacial resolution (information preserving) are discussed also.

Pseudorandom sampling pattern and image digitization have been investigated as a technique to reduce the video bandwidth before transmission. Most techniques commonly used involve some sort of transform which makes transmission in real time impossible. In order to transmit in real time, video quality may be affected in one or more of the following ways - possible loss of data during compression; reduction in sharpness of the picture when played back and the possibility of having unwanted video effects such as dot crawling and dot patterning. Hughes Aircraft Company has conducted a study on these video effects versus different sampling techniques and part of the system design is based on their findings. The pseudorandom pattern sampling and image digitization was investigated for the NASA Johnson Space Center as a part of the Space Station development activities. The need for this development is to avoid a signal overload condition in the multiple access communication link. All downlink signals will be assigned levels of priority before transmission and video signals are usually assigned a lower priority. Therefore, the bandwidth available for video transmission will be limited. This creates the situation where an image must be compressed before downlink. The advantage of using the pseudorandom sampling pattern and image digitization technique is that real time information can be obtained and the viewability of the displayed picture can be maximized even in the high compression ratio mode. This report focuses on two main areas the hardware design/implementation and the sampling technique used for the image compression.

In conventional constant transmission rate image coders the image quality is varied in dependence on the current activity in the image sequence in order to obtain a constant data rate. Flexible transmission rate will be available in the future along with packetized networks and a constant image quality could be achieved. Using a 2 Mbit/s constant transmission rate codec for video conference we have studied three modifications of this coder to adapt it to variable transmission rate image coding: open loop, closed loop and constant MSE distortion. It is shown that none of these modes yields satisfying results neither from the image quality nor from the channel rate requirements point of view. After a thorough discussion of the pros and cons of above modifications the concept of a combined quality and bit rate controller is introduced. It is capable of satisfying both the viewer by a constant perceived image quality and the network by suitable coder output bit rate statistics. For a constant image quality a quality measure was developed reflecting the perceived image quality very well. Using this quality measure we are able to adapt quantization to the visibility of errors and thus prevent waste of bits.

A modular hardware architecture for video coding at p x 64kbit/s data rates is described. The codec uses several digital signal processors (DSPs) and can be viewed as a single instruction multiple data (SIMD) computing architecture. Every image in a sequence is divided in regions of horizontal strips and each region is operated by its an processor. These local processors communicate with a central processor which codes (decodes) the cosine transformed frame differences. Lateral communication between adjacent processors is also permitted. This is done by memory sharing and allows comparisons between blocks situated in neighbouring regions, as required by most motion estimation algorithms. The codec is built using the modern TMS320C30 digital signal processor. The number of processors used in both the coder and the decoder depends on the application. This is a consequence of the modular design and allays the machine to be configured to suit a particular algorithm complexity or a desired quality of the coded image.

The RMS-1000 Reconfigurable Mobile System is a highly mobile ground station with direct application to the needs of users requiring real-time access to information gathered by reconnaissance platforms. The system combines high-performance computing and image processing workstations with command, control, and communications capabilities in a lightweight self-contained shelter. The RMS-1000 provides a capability for receiving and processing many forms of intelligence data, including imagery. Background information is presented which shows the RMS-1000 as the culmination of over 38 years of tactical ground station development. Traditional methods of reconnaissance exploitation are compared with those used in the RMS-1000.

A system is proposed for the high precision on-line verification of the minia-ture spring structure, including overall height, diameters of various coils as well as pitches between neighboring coils of the miniature conical springs. High preci-sion measurements without physical contact and short processing time are achieved. Deformations of any kind on the conical springs can be identified even from the worst viewing direction.

An expression of images called "Tree-Structured Figure Expression" is here proposed.' The data structure and the algorithm for transformation from digital image to digital figures is described. A new image processing system architecture has been realized for a super mini computer. Procedures by which two kinds of algorithms, digital-image based algorithms and digital-figure based algorithms, can be effectively combined are demonstrated. This new architecture brings higher performance and higher expansibility to image processing systems.

Detection of scratches and cracks is a key problem in the quality assurance in crystal production. difficulties in these inspections are due to the fact that such defects are so thin and short that they are almost indiscernible from the textural background of the crystal unless a special illumination and torsional viewing mechanism are provided. The purpose of this paper is to propose an effective method to automatically segment a microscopic image into separate textural structures, and eventually delineate on the quartz blanks their boundaries which show the locations of the cracks and also the scratches.

Numerous computations are required for acquiring and tracking multiple objects in real time. This paper discusses implications for architectures and bandwidths when video information is used with Area Parameter Analysis for tracking.

This paper discusses the implementation of an Image preprocessing technique. Suitable hardware and software are developed and tested to match two images of the same scene, obtained by two different sensors. Matching is required to carry out correlation techniques. The data corresponding to the required reference image will be taken from a RAM area and preprocessed image data will be sent to correlation tracker by RS 232C communication interface.

The ideal image processing system would be one that people already know how to use. Its components would be interchangeable with other computers in the office and in the factory. It could be quickly programmed to do its job by non-programmers and non-typists. And it would communicate and print reports over the same network used by all the other computers in your company. Does this sound like science fiction? It's not. The image processing system in this scenario represents a trend toward standardization in the industry. Image processing suppliers are realizing that they must move away from their existing proprietary architectures and customized software systems in favor of standard computers having a wide base of support hardware and software, and a wider pool of knowledgeable users.

This paper describes a polynomial approach to the algebraic representation of binary and gray images for machine vision. Most of the image processings can be done by the template polynomial operations. The advantage of the polynomial approach over the classical methodology is discussed.

In this paper, a stop sign recognition system is viewed from a morphological filtered response in the log-polar domain. This tessellation makes it feasible to predict and calculate closure rates for a visually controlled autonomous land vehicle. Development of the recognition system for octagonal shape and its subsequent mapping to the log domain are investigated. Simulation and prediction of a vehicle's closure rates as a function of angle of view and velocity are compared to actual video when a vehicle is approaching an intersection. The impact of this research on vehicle navigation and guidance is given. A polar exponential grid sensor mapped to the log domain possesses subtle and powerful properties which may be applied in mobile applications. It is concluded that mapping to the log conformal domain makes optical flow a manageable problem in dealing with vehicle motion.

Spread Spectrum System (SSS) or Code Division Multiple Access System (CDMAS) has been studied for a long time, but most of the attention was focused on the transmission problems. In this paper, we study the results when the code division technique is applied to the image at the source stage. The idea is to convolve the N different images with the corresponding m-sequence to obtain the encrypted image. The superimposed image (summation of the encrypted images) is then stored or transmitted. The benefit of this is that no one knows what is stored or transmitted unless the m-sequence is known. The recovery of the original image is recovered by correlating the superimposed image with corresponding m-sequence. Two cases are studied in this paper. First, the 2-D image is treated as a long 1-D vector and the m-sequence is employed to obtained the results. Secondly, the 2-D quasi m-array is proposed and used for the code division multiplexing. It is showed that quasi m-array is faster when the image size is 256x256. The important features of the proposed technique are not only the image security but also the data compactness. The compression ratio depends on how many images are superimposed.

An image registration algorithm combining image analysis and signal processing techniques is applied to a sequence of complex images. The frequency-domain registration technique described here is faster than the conventional spatial domain correlation techniques and does not suffer from the non-uniqueness of solution. The specific technique used in this work involves power spectrum for rotation correction and power cepstrum for translation correction. Furthermore, a comparison of another frequency-domain translation correction technique namely phase correlation and power cepstrum technique reveals clear superiority of the latter method for noisy images and for specific application namely registration of sequential complex images. The general procedure of employing feature extraction techniques prior to registration for simplifying the image and subsequent registration of images based on power spectrum and 2-D cepstrum techniques provide an unambiguous, accurate and fast method of comparison for a broad range of sequential complex images.

The zero-crossing test of the second directional derivative is regarded by many image processing researchers as the optimal method of edge detection. Fast implementation of a zero-crossing detector for an image that has been operated upon by a Laplacian of Gaussian (LOG) convolution is difficult; execution speed is, in fact, the primary disadvantage of edge detection by LOG. Another drawback to this method is the difficulty in programming the algorithm itself. Use of efficient library functions to perform the required LOG convolutions and tradeoffs between multiple cascaded filters and a single filter with a spatially large impulse response will be examined. Finally, a convolution and look-up table-based implementation of a zero-crossing detector will be explained.

The positions of edges along each row and column of an NxN image are found by computing the singular value decomposition of a matrix formed with the 2N point 1-D Fast Fourier transform of a symmetric signal corresponding to the row or column and locating the peaks in the inverse of the sum of the absolute values of the 2N point Fast Fourier transforms of the left singular vectors corresponding to the group of smallest singular values of this matrix. The computational complexity of the proposed procedure is essentially 0(1V2log(N)).

This contribution presents an efficient method to estimate the motion parameters of rigid objects with curved. sufaces. The problem is formulated as a model-based estimation problem which is solved by an iterative minimization strategy based on ordinary grey-scale images. Therefore no a priori knowledge like corresponding points or the displacement vector field is required. The algorithm also can be used to segment a more complicated scene into different moving objects, whose motion parameters are estimated. As an application the segmentation and motion estimation in the context of low bitrate video coding is discussed.

A highly parallel technique for linking line segments into curved contours or boundaries has been developed. This bottom-up organization of the line segments utilizes information in the line data itself and perceptual organization rules which are domain independent. The line segments can be detected on a local basis and are amenable to parallel computation.1 A line pyramid structure is employed where the two bottom levels of the pyramid hold the line segments extracted from predefined image windows (e.g. 4 x4 and 8x8). The pyramid is based on a 4 x4 overlapping neighbourhood and each processing element is connected to four parents and sixteen children. Starting from the lower level, each processing element performs grouping on the line data supplied by all of its children but it only keeps the grouped line segments supplied by its central 2x2 children for passing on to the parents. The main grouping criteria employed to achieve the aggregation of line segments are proximity, Mares theta-aggregation,2 curvilinearity, continuity and similarity in line contrast. By the time the root processor of the pyramid is reached, all the groups of line segments are formed and they are then replaced by contours. The main features of the method are its parallel implementation on a pyramid architecture computer and use of domain independent perceptual grouping principles.

How to use one-chip processor and small size memory to recognize the shape of object is the most important topic in modern industrial application. Usually, it is impossible to get the front view of fast flying object, the images taken at the view direction are called plane projection images. The plane projection image is composed of "object" and "background", that is what we say pixel recursion image The image processing by the using of pixel recursion is the most simple and fast approach. The pixel recursion image can be achieved by the scanning, and the image after size and orientation normalization can be recognized by the Quad String Array (QSA) approach. The size normalization can be achieved by the scanning device and the orientation normalization can be handle by the Circular Layer String method. In the orientation normalization, the principle axis of unknown image is achieved by rotation and compared with the standard patterns which are pre-estimation. The unknown image can divided into several subimages and each subimage is coded by the Vertical String Array. By the using of Vertical String Array, the shape of unknown object can be described precisely. In this paper, the experiment results of Vertical String Array, Random Walk and Gauss-Markov-R-process are also discussed. Index Terms--Pattern Recognition, String Array Code, Circular Layer Code, Image Processing, Kolmogorov Complexity Program, Turing Machine, Automaton, Quad String Array, Gauss-Markov-R-process, Random Walk, Pixel recursion, Block matching Algorithm.

A new image processing technique has been developed which uses an operational eigenvector associated with the null eigenvalue (e.i.o) together with supplementary information to limit non-uniqueness associated with standard methods. A field of e.i.o. corrections is iteratively imposed to minimize the test function(s) reflecting solution agreement with the supplementary information. Computer simulation with supplementary information consisting solely of the global source field strength second moment value results in dramatic improvement over standard methods.

We present some preliminary results of a study aimed to assess the actual effectiveness of fractal theory in the area of medical image analysis for texture description. The specific goal of this research is to utilize "fractal dimension" to discriminate between normal and cancerous human cells. In particular, we have considered four types of cells namely, breast, bronchial, ovarian and uterine. A method based on fractal Brownian motion theory is employed to compute the "fractal dimension" of cells. Experiments with real images reveal that the range of scales over which the cells exhibit fractal property can be used as the discriminatory feature to identify cancerous cells.

In this paper the radial digital image scan is considered. The main result of the paper is the number of radial digital lines needed to cover the image composed of pR x pR pixels, which is pR Application of such scan to detect defects on a silicon wafer is presented.

In order to take a promising Automatic Target Recognition (ATR) algorithm or system from the development stage to the point where it is robust enough for the application which it is intended for, a considerable amount of testing, training, and evaluation is required. This can take years. For example with seemingly slight distortion in the target/clutter characteristics, the algorithm behavior becomes erratic and random. In addition, no well defined method currently exists for designing image processing algorithms. Sophisticated and intelligent tools are needed in order to perform such a task in a rapid & efficient manner. A prototype system has been developed which uses a blackboard architecture and allows rapid mapping of procedural algorithms into a knowledge based development environment. This system was developed at Honeywell Systems and Research Center. Coupled with the graphics interface, this provides a powerful tool for prototyping & tuning ATR algorithms. Significant improvements were obtained from the knowledge based version over the conventionally coded procedural version.

A technique for designing digital filters that preserve edge position during image resampling is described. The technique consists of designing a set of separable filters, with each filter designed as an all pass filter with nonlinear phase and a set constrained amplitude. The technique is compared to existing techniques.

An optical implementation of a new class of nonlinear filters known as the finite impulse response - median (FIR-M) hybrid filters is proposed. These filters preserve both details and edges better and are computationally more efficient than conventional median filters.

We are developing multimedia information systems, concerning images, figures, characters and numeric values. In this paper, we describe a similarity retrieval method, that retrieves images or figures similar to given key images or figures. In this retrieval method, the topological structure of the key and that of the stored data are utilized to determine the similarity. For this determination, a level of similarity is designated by a user. According to this level, various traits of the stored image contents are retrieved (for example, whether the orientation of each element is concerned with the similarity or not). This retrieval method was applied to a house plan database and to a piping and instrumentation diagram database.

In image analysis, most problems arise from the segmentation process. These problems notably increase when one has to work on SAR images, due to the speckle noise affecting this kind of data. A large number of filters have been proposed in order to reduce speckle, but results are not satisfactory in all cases. The idea suggested in this paper is to by-pass the speckle drawback by performing segmentation not only on physical data but also on virtual ones. The latter are derived from the former through some processing step (e.g., textural analysis). Segmentation is then achieved by analyzing these different data simultaneously, thus recovering uncertain situations. Some results are reported and discussed: they evidence the high performances allowed by this approach, even when speckle noise is not reduced.

This paper proposes a procedure to investigate whether the statistical independence among the pixel images of a picture affects the statistical image segmentation. The results of investi-gation shows that ignoring this independence will degrade the quality of the segmented images: the more pixels which belong to the major image regions will be misclassified into the minor image regions, an artifact of statistical image segmentation. The procedure utilized in this in-vestigation can be used as an accurate statistical image segmentation technique .

A contour transfer device has been developed to project a computer designed shape onto the surface of a radiotherapy patient. The device consists of a He-Ne laser, an optical system with a 2-D scanner and a scanner driver with microprocessor control. The X-Y scanning is accomplished by using a twin-mirror deflector; the scanning pattern is a function of the rotated angles of the mirrors. A program is proposed to calibrate the device and interpret the computerized contour so that the drawing pattern of the device can precisely reproduce the computer data. The device we developed is a versatile instrument with a number of applications in auto-drawing areas. It is to be used to project the intended treatment field of a radiotherapy patient onto the patient. This is the critical last step needed to utilize a computerized tomography (CT) scanner for radiotherapy treatment planning. The device is constructed in a compact fashion and will be mounted on the gantry of the CT scanner with the following technical parameters: (1) maximum scanning field is 30 cm by 30 cm on a 20-25 cm thick patient, (2) frame rate is greater than 25 frames/S, and (3) random error is within + 1 mm.

The template matching method has been more successful in character symbols recognition than in object recognition. The major disadvantage in the later type of problems is that the number of templates that needs to be stored increases with the number of different objects to be recognized. The relatively small and fixed number of basic symbols used in writing any word in a given language is what makes template matching the most widely used method in problems related to the recognition of handwritten or printed character symbols.

An object illuminated by a source produces a scattered signal; this signal depends upon both the source and the physical properties of the object. The problem of deducing coordinates, shape and/or certain physical properties of the object from the measurements of the returned signal is an inverse problem called echo imaging. The problem of echo imaging arises in medical imaging, remote sensing (radar; sonar; geophysical exploration), and non-destructive testing. In this paper, we address the problem of imaging an object form its returned signals using a phased array. Our approach is to exploit the array's various radiation patterns and the recordable portion of the returned signal's spectrum to generate the data base for this echo imaging system. Rapid steering of a phased array's radiation patterns can be achieved electronically. These steered waves can be utilized to synthesize waves with varying angles of propagation. In this case, the recorded returned signal for each direction of propagation can be viewed as data obtained by a bistatic array configuration. We first formulate the imaging problem for a plane wave source in a bistatic configuration. We utilize the two-way propagation time and amplitude of the returned signal to relate the object's properties, reflectivity function and coordinates, to the measured data (system modeling). This relationship is the basis for deducing the object's reflectivity function from the recorded data (inverse problem). We then extend these results for an arbitrary radiation pattern and synthesized radiation patterns generated by an array capable of beam steering in cross-range. We show that the recorded returned signals can be related to the spatial frequency contents of the reflectivity function. We also show that these array processing principles can be utilized to formulate a system model and inversion for synthetic aperture radar (SAR) imaging that incorporates wavefront curvature.

Size distributions of particles within a binary image can be generated by morphological filtering processes known as granulometries. These filter the image by structuring elements of ever increasing size, the ultimate result being a distribution correlated to particle size. The granularity of electrophotographic solid area images can be analyzed by applying morphological granulometries that generate size distributions providing greater information than that currently being used for feedback control systems. This is accomplished by examining the microstructure of the image. The approach discussed in the present paper employs simulation techniques employing existing magnetic brush development and optical density transform models to yield particle area distributions in electrophotographic images. The application of a granulometry to a simulated particle area distribution produces a morphological size distribution whose sample statistics are characteristic of various parameters in the electrophotographic process.

A fast implementation of the discrete Fourier transform (DFT) for spectral analysis of two-dimensional images is presented. The implementation time is order N, where N is the width of the image. The implementation is based on an old algorithm and a new architecture. The algorithm is the Chirp Z transform (CZT) implementation of the DFT. The architecture is a multiprocessor, mesh-connected, single-instruction, multiple-data (SIMD) scheme based on the Martin Marietta Geometric Arithmetic Parallel Processor (GAPPTM) chip. Using this implementation, image texture may be adaptively filtered for real-time image understanding applications.

Our newly developed video weld authentication system recognizes common structural characteristics between welds. A weld fingerprint is archived for comparison at a later time to the reference weld for verification purposes. This verification procedure is performed using an IBM-PC/AT and state-of-the-art PC image-processing hardware and software.

This work derives the frequency domain representation of an image produced by a scene moving under costant acceleration and non-zero initial velocity. The effects of the acceleration are contrasted against those of the velocity, as they lead to remarkably different scenarios.

Reconstructed MRI images must ideally be real and positive, since they correspond to density distributions of objects, which by definition are real, positive quantities. In practice however, most reconstructed images become complex due to a shift in data from the nominal origin in k space. In some situations such as constant readout gradient, these phase shifts do not affect the magnitude of the reconstructed images, and can therefore be easily determined and eliminated. However, when the readout gradient is sinusoidal and the frequency plane is scanned with data reversal on alternate lines, time delays between the start of data acquisition and the start of the readout pulse become different for even and odd lines, and result in a ghost separated by half the image size. In this paper, we describe ghost cancellation algorithms for restoration of MRI images in medical applications. Our approach is to model the effect of the time delays and the asymmetry of the sinusoidal readout gradient for even and odd lines by two phase functions relating the actual object density to even and odd parts of the observed image. We then exploit a priori information about the phase functions in order to estimate the true object density. Examples of application of this ghost cancellation approach to liver and heart images will be presented.

The nineteen nineties will usher in a new era in image processing systems. The maturation of the technologies required to capture, store, process, retrieve, display and distribute documents electronically will enable government and industry to implement large scale, highly productive and cost effective, document image processing (DIP) systems. This new information processing technology will greatly enhance our ability to deal with the ever increasing volume of paper documents that have become the hallmark of our knowledge based societies. This paper briefly discusses the forces driving the evolution of DIP systems and the DIP marketplace. Technical considerations and technology trends, involved in determining a realistic and successful approach to implementing a DIP system, are also presented.

A proof-of-concept system has been designed to detect airplanes in aerial imagery using their general shape. The system has three distinct levels of processing: 1) a low level that extracts line segments from raw image data, 2) a middle level that combines the line segments into geometric feature primitives, and 3) a top level that checks for positional coincidences of features to perform the airplane detections. The system has been tested on a limited set of images with encouraging results. More extensive testing of an enhanced version of the system is imminent. The initial testing seems to show that geometric features can be efficiently used to detect airplanes with low false alarm rates. The detection process is relatively insensitive to view angle, object orientation, and sun angle. No a priori scene knowledge is used. Furthermore, low contrast conditions and limited occlusion can be tolerated. The airplane description provided to the system to drive the detection process is a very general one, suitable for a broad range of look-down views. The speed performance of the system is also good. A low level line segment extraction algorithm has been invented to greatly improve the speed and reliability of the initial image processing steps compared with similar older methods. The middle level of the system has also been designed to give efficient performance and relatively fast execution speeds. The overall speed results are sufficiently good to suggest the future possibility of near frame speed performance, given a suitable custom hardware architecture.

Objects that vibrate at frequencies higher than 1.0 MHz usually have vibration amplitudes far below the first fringe minimum (about λ/5) of holographic detection. To get a complete phase and amplitude map of such high frequency vibrations we have combined phase modulation with phase shift techniques, using an electro-optic modulator and a digital image processor. Results showing the vibrations of a ceramic crystal (3.8 MHz), with a complete phase and amplitude map, are presented.

This work is focalized on image segmentation for aerial features extraction for image based aircraft navigation. The basic concept of the method is the determination of frontiers of objects by using correction of local histograms. The complete segmentation scheme includes a region growing approach based on correlation test.

A method is shown in this paper to calculate the possible character recognition error rate, which is originated from the optical transfer functions and sampling performance of the scanner. The increase of the possible reading error is the result of the image degradation through the optical transfer, and can be measured by the so-called similarity error. Theoretical and experimental results are in a good accordance. The different character types, scanners and printer outputs can be characterized by this method, as well.

This paper presents a new special computer vision system to inspect the defects in welding seams of acetylene cans based on X-ray pictures. The system configuration is introduced briefly and basic problems are discussed. A high resolution CCD linear array camera is used for scanning pictures displaied the screen on X-ray image intensifier. We propose several new methods to determine the defect position and representation, to reconstruct the defect boundary, and to extract local features for defect classification. Experimental results show that the proposed methods are effective. The system can distinguish a minimum of l6 decimillimetres defective patterns. Further investigations are suggestedd on more generality of the methods and more reliable stability of the inspection.

Morphological processing operations are necessary for many difficult automatic target recognition (ATR) problems. All of the basic morphological functions can be performed on an optical correlator. This paper emphasizes the hit-or miss morphological transform and how it provides improved part recognition.

Historically, computer graphics and image processing technologies and applications have been distinct, both in their research communities and in their hardware and software product suppliers. Computer graphics deals with synthesized visual depictions of outputs from computer models*, whereas image processing (and analysis) deals with computational operations on input data from "imaging sensors"**. Furthermore, the fundamental storage and computational aspects of these two fields are different from one another. For example, many computer graphics applications store data using vector formats whereas image processing applications generally use raster formats. Computer graphics applications may involve polygonal representations, floating point operations, and mathematical models of physical phenomena such as lighting conditions, surface reflecting properties, etc. Image processing applications may involve pixel operations, fixed point representations, global operations (e.g. image rotations), and nonlinear signal processing algorithms.

The proposed approach to stereo image coding takes advantage of the singleness of vision property of the human visual system. Experiments show that a stereo image pair, in which one of the images is low-pass filtered and subsampled, is perceived as a sharp 3-D image. The depth information is perceived due to the stereopsis effect, and the sharpness is maintained due to the details in the non-filtered image. A methodology for the evaluation of the compression effects on the 3-D perception of stereo images is presented. It is based on measurements of response-time and accuracy of human subjects performing simple 3-D perception tasks.

In this study, a technique which enhances image contrast using differentials of radiances being measured from successively over-lapping footprints is discussed. One of the immediate advantages of this approach is to improve spatial resolution when the practical size limitation of detectors is reached. When diffraction is considered, radiometric accuracy at the higher spatial frequencies is decreased even though edge response (boundary crossing sharpness) will improve. For instance, an imager system based on this principle could have a primary aperture no larger than 30 cm diameter yet be capable of resolving a ground footprint as small as 2 m from a 780 km altitude.

This paper describes a method of precisely aligning two digitized images prior to a point-by-point comparison of the images. Histogram specification to a reference is used to align the images in the photometric axis. A novel method of estimating the location of the cross-correlation peak using a parabolic surface is given. The accuracy of spatial registration using this method is 1/16 pixel. A hardware implementation and error analysis is also included.

"Real time" surgical treatment planning utilizing multimodality imaging (CT, MRI, DA) has been developed to provide the neurosurgeon with 2D multiplanar and 3D views of a patient's lesion for stereotactic planning. Both diagnostic and therapeutic stereotactic procedures have been implemented utilizing workstation (SUN 1/10) and specially developed software and hardware (developed in collaboration with TOMO Medical Imaging Technology, Southfield, MI). This provides complete 3D and 2D free-tilt views as part of the system instrumentation. The 2D Multiplanar includes reformatted sagittal, coronal, paraaxial and free tilt oblique vectors at any arbitrary plane of the patient's lesion. The 3D includes features for extracting a view of the target volume localized by a process including steps of automatic segmentation, thresholding, and/or boundary detection with 3D display of the volumes of interest. The system also includes the capability of interactive playback of reconstructed 3D movies, which can be viewed at any hospital network having compatible software on strategical locations or at remote sites through data transmission and record documentation by image printers. Both 2D and 3D menus include real time stereotactic coordinate measurements and trajectory definition capabilities as well as statistical functions for computing distances, angles, areas, and volumes. A combined interactive 3D-2D multiplanar menu allows simultaneous display of selected trajectory, final optimization, and multiformat 2D display of free-tilt reformatted images perpendicular to selected trajectory of the entire target volume.

An algorithm for three dimensional identification of the coronary artery by biplane angiogram is described. The algorithm is based on a few assumptions of the anatomical structure of the coronary artery and it succeeds in reconstructing the 3-D structure almost automatically using information such as diameter of the artery and the direction of it at node points. The results are displayed on a CRT as a stereo pair for binocular vision or as a rotating 3-D structure. Combining these results with a 3-D computer display of a myocardial scintigram, an effective tool for evaluating or understanding mutual image information is obtained. For the myocardium, superimposition with a conceptual coronary artery model is also effective. Experiments with clinical data show the usefulness of this procedure for understanding the degree and region of ischemia.

This paper describes both method and first results of an indoor / outdoor 3D reconstruction application. Images are supposed to come from man-made environments. The method is based on a new cooperative segmentation designed from a rough model of such scenes. This modelisation consists of classically representing scenes by a universe composed of polyhedra, that we call general model. This allows to deduce an operating model, constituted by all the typical 3D tokens of a polyhedron: edges, vertices, plane faces. Their projections are 2D primitives as edge segments, junctions of segments, regions. Then a graph groups these features together . The segmentation principle is based on a loop of two detectors of dual image properties (homogeneity and disparity): each of them is controlled by the results of the other.

Linear programming has been used to solve a filter design problem typical of video applications: the trading-off of frequency domain performance versus space-time performance. The technique has proven useful to design frequency selective filters without ringing used in PAL decoding and scanning rate conversion systems.

This paper describes an application-specific engineering workstation designed and developed to analyze imagery sequences from a variety of sources. The system combines the software and hardware environment of the modern graphic-oriented workstations with the digital image acquisition, processing and display techniques. The objective is to achieve automation and high throughput for many data reduction tasks involving metric studies of image sequences. The applications of such an automated data reduction tool include analysis of the trajectory and attitude of aircraft, missile, stores and other flying objects in various flight regimes including launch and separation as well as regular flight maneuvers. The workstation can also be used in an on-line or off-line mode to study three-dimensional motion of aircraft models in simulated flight conditions such as wind tunnels. The system's key features are: 1) Acquisition and storage of image sequences by digitizing real-time video or frames from a film strip; 2) computer-controlled movie loop playback, slow motion and freeze frame display combined with digital image sharpening, noise reduction, contrast enhancement and interactive image magnification; 3) multiple leading edge tracking in addition to object centroids at up to 60 fields per second from both live input video or a stored image sequence; 4) automatic and manual field-of-view and spatial calibration; 5) image sequence data base generation and management, including the measurement data products; 6) off-line analysis software for trajectory plotting and statistical analysis; 7) model-based estimation and tracking of object attitude angles; and 8) interface to a variety of video players and film transport sub-systems.

This paper reports a model-based system for recognition of 3D overlapping convex objects with planar and curved surfaces. We use photometric stereo to generate a needle map, segment the needle map into planar and curved surface patches and represent a scene by means of an attributed surface patch graph. A corresponding attributed surface patch graph is used as a model for each prototype object to be recognized. The models are automatically generated from data structures of a commercial CAD-system. Our approach to recognition consists of generation of feasible interpretations and EGI matching. Local geometric constraints are used to prune the search space of feasible interpretations. The recognition system has been completely implemented. Experiments are shown for scenes of touching and overlapping objects with planar and curved surfaces.

The block segmentation and block classification of digitized printed documents segmented into region of texts, graphics, tables and images are very important in document analysis and understanding. Conventionally, the Constrained Run Length Algorithm (CRLA) has been proposed to segment digited document, but failure may occur due to improper constraints. Especially, it usually leads to failure about block segmentation when the documents are complicated and inclined. They could only deal with the text part for block classification without any certain rules, and couldn't succeed in effective classification and even lead to wrong classification. In this paper, a powerful approach for document analysis named "Automatic Local Sequential Segmentation and Hierarchical classification" is proposed. Our results show that this algorithm is an efficient approach for block segmentation and block classification.

A method of three-dimensional reconstruction of coronary artery configuration from biplane angiographic images is dis-cussed in this paper. The proposed method is based on the geonetric mathematics of X-ray imaging apparatus in cineangiogram and the tracking of leading edge of injected contrast material into each vessel for identification of corresponding points on two images taken from orthogonal views. The three-dimensional morphology of coronary artery can be obtained by this recon-struction procedure. The reconstructed arterial configuration is displayed as a shaded surface model at any arbitrary angle for viewing. Analysis of reconstructed arteries provides the useful morphologic and functional information on vascular system for diagnosis of vascular diseases.

A new type of interactive color quantization method to realize both high speed and high S/N ratio quantization, and flexible color selection reflecting designer's feeling, is proposed in this paper. In the proposed method, flex-ible clustering of color points in the 3-D RGB space is performed using two selection parameters, (D), which corresponds to a maximum distance allowed between a selected representative point and color pixels and (I), which indicates 'importance' of arbitrary given rectan-gular areas. Clustering is achieved by high-speed re-trieving and deleting of color points in the RGB space within a D-sphere centered at a representative point. A three-dimensional pattern data structure named BD-tree can effectively employed to realize this high-speed retrieval. With smaller D-value, impressive color even in the small area can be preserved. With larger D-value, smooth color change in the large area can be expressed. With larger I-value, the corresponding area is empha-sized and relatively more colors are assigned to the area. Many experimental results will be shown in the paper, which reveal the superiority and the flexibility of the proposed method.

Since the determination of the homologue points between two stereo images is the most serious problem in stereo vision, many different systems have been studied to solve it, using some of them auxiliar equipment such as laser beams. This problem can be drastically reduced when those images are previously preprocessed and most of the redundant information contained on them is eliminated. For a great number of scenes the contour of the image contains information enough to be analyzed. For many applications this information can be further reduced keeping only the straight line edges of their contours, maintaining yet information enough to be interpreted. Having these features of the images the search for the correspondence of some of the points of the edges is greatly reduced and thus it is possible to develop a system that can extract a 3D map of the environment in real time. In this paper a system able to generate a 3D map of the main parts of many scenes, the edges, is presented.