In the design of any high quality scanner utilizing a linear CCD imager, a set of features is required that is largely independent of the application. At the front end, sensor clocking signals are generated and analog-to-digital conversion occurs. Next, corrections for differences in pixel dark current and gain are applied, along with pixel and/or line summing to improve the signal-to-noise ratio. The corrections and summing can be done by software, but significant speed improvements are obtained when a hardware solution is employed. The log of the data may be calculated, and the data must be supplied to some form of image storage device. RAM memory on a standard computer bus, such as VME, provides a straightforward means to configure an image storage, so long as a means to transfer sensor data to the memory is provided. This paper will begin with a review of the various ways in which data from a CCD scanner is processed for use in a digital imaging system, including the correction for dark noise variations and gain errors, data space transformations, and the use of pixel or line summing to improve signal-to-noise ratio. Following that, a high speed hardware implementation of these functions will be described.

This paper presents a film scanner that utilizes linear CCD array technology and adjacent pixel summing of m pixels to scan large format films. Pixel summing is employed to optimally increase the dynamic range of the system, providing the ability to scan films with wider exposure latitude without compromising radiometric resolution. The digitized CCD data is presented to a hardware processing unit that acts as both an interface to the data storage device and controller of clocks on the CCD array. If the first-of-m data level is below a predetermined threshold, the CCD array clocks are modified to allow the next m-1 charges to accumulate in the output stage of the CCD array (analog summing). If the first-of-m data level is above the predetermined threshold, the processing unit switches to digital summing and the next m-1 pixel values are digitally added to the first-of-m value. This pixel summing scheme not only increases SNR and dynamic range of the scanner, it also allows the user to change m, the number of pixels summed, giving him the ability to vary scanner spatial resolution without changing the optics.

We have developed a video rate color image processing system "Color-IDATEN" that can process time-varying color images at video rate. In this paper, we present the system configuration, the new processing modules, and the experimental results. The system consists of a host CPU, a network unit, and three monochrome image processing units (IPUs) with three new processing modules: dynamic masking, color processing, and feature extraction. These modules are controlled by the host CPU and can process each pixel of images from the network in 100 ns. The results from the experiments with the time-varying color images showed that Color-IDATEN effectively erased and extracted moving objects and particular colored areas.

Direct color mapping method using 3-dimensional look-up table is promising for high accuracy and high speed color correction, instead of matrix masking method. But it needs large storage capacity which leads to cost increase. So various ideas using interpolation technique were proposed for reducing the storage capacity. This time, two new ideas for reducing the storage capacity are introduced and evaluated. One is to eliminate the part of 3-dimensional look-up table memory corresponding to the outside of color gamut of output device. Its reduction ratio is about 25%. The other is to reduce the look-up tables instead of multiplier for 3 X 3 matrix interpolation. Its reduction ratio is about 20% in the case of 8 bit X 3 input and 8 bit X 3 output using lower 3 bit interpolation. By the combination of these ideas, the storage capacity can be reduced down to 100k byte in that case.

This paper discusses a technique for video enhancement averaging of the surrounding area of each sampled cell when the source of the electronic image is the output of a CCPD linear array. In order to achieve the high speed output required for imaging from a document moving at up to 400 inches/second with resolution of up to 240 pixels per inch, it is necessary to sample in eight parallel segments of 128 cells and reconstruct these parallel outputs as if they were coming from one continuous 1024 linear source. This requires that one provide a means of bit averaging the background surrounding each sample and specifically provides a means for "healing" the seam between the last cell of each of the eight segments and the first cell of the next. All eight segments present cell #1 during the same clock time and present the next cell during the next clock time; therefore, cell #128 of segment #1 is "looking" at the adjacent vertical position to cell #1 of segment #2, but segment #2, cell #1 was clocked out at the same time as cell #1 of segment #1 or 128 clock times earlier. This makes it necessary to provide a means of "healing" the seams between segments thus allowing the correlation of each cell with the surrounding area that includes cells sampled from adjacent segments. This segment to segment data passing is necessary between the first and last of several cells per segment. The exact number depends on the area size of the correlation desired. The area used to average and correlate is usually between 3x3 and 7x7 in order to obtain optimum center cell comparison. The area size must be odd in order to provide the same number of background cells above and below or to the right and to the left of the center cell.

XVision is an example of a comprehensive software system dedicated to the processing of multidimensional scientific data. Because it is comprehensive it is necessarily complex. This design complexity is dealt with by considering XVision as nine overlapping software systems, their components and the required standards. The complexity seen by a user of XVision is minimized by the different interfaces providing access to the image processing routines as well as an interface to ease the incorporation of new routines. The XVision project has stressed the importance of having; 1) interfaces to accommodate users with differing preferences and backgrounds and 2) tools to support the programmer and the scientist. The result is a system that provides a framework for building a powerful research, education and development tool.

A comprehensive approach to estimation is presented that integrates Kalman estimation and minimum mean square error (MSE) linear predictive transform (LPT) signal source modeling. The approach is strictly optimum when the signal source is a linear recursive system of the Kalman type and the channel contributes additive white noise. The technique is an improvement over classical Kalman estimation for two fundamental reasons. First, it directly addresses the signal source modeling problem of Kalman estimation and, secondly, it provides an inherent transformation mechanism that allows for simplifications in the design and implementation of the estimator. These ideas are illustrated with noisy monochrome images, where it is found that simple and robust LPT estimators ( filters or smoothers) with significant signal to noise ratio (SNR) enhancement can be obtained. In particular, it is shown that when applied to images, LPT estimators are easier to design and implement, without any SNR degradation, than classical Kalman estimators by a factor that approaches four in the smoothing case. Further simplifications can be achieved with a negligible loss in SNR performance by using approximations which are suggested by inherent transformation and robustness properties of the LPT estimators. Numerous areas for further investigation are also discussed. In particular, a comprehensive approach to control is developed which integrates minimum MSE LPT signal source modeling and linear quadratic white (LQW) control, of which linear quadratic gaussian (LQG) control is a special case. The control approach is strictly optimum when the signal source is a linear recursive system of the LQW control type and the channel contributes additive white noise.

Hierarchical classifiers are discussed and a new implementation in digital image processors with a pipeline architecture is presented. This implementation makes optimal use of the different layers of look up tables that generally exist in all machines with this architecture. Classification can be considered to occur at almost real time, as far as the unit structure of the tree associated to the classifier is concerned. As the hierarchical classifier makes optimal use of all the information available, fast decision rules can be applied in most of the nodes, thus reducing considerably the overall computational burden. An example is provided in the context of classification of multi-temporal remote sensed data. To apply the technique, the evaluation of the features must have been done previously.

Many image transmission/storage applications requiring some form of data compression additionally require that the decoded image be an exact replica of the original. Lossless image coding algorithms meet this requirement by generating a decoded image that is numerically identical to the original. Several lossless coding techniques are modifications of well-known lossy schemes, whereas others are new. Traditional Markov-based models and newer arithmetic coding techniques are applied to predictive coding, bit plane processing, and lossy plus residual coding. Generally speaking, the compression ratio offered by these techniques are in the area of 1.6:1 to 3:1 for 8-bit pictorial images. Compression ratios for 12-bit radiological images approach 3:1, as these images have less detailed structure, and hence, their higher pel correlation leads to a greater removal of image redundancy.

Members of the International. Standards Organization ISO/IEC JTC1/SC2 Working Group 8 (Coded Representation of Picture and Audio Information) and the Consultative Committee of the International Telegraph and Telephone (CCITT) Study Group VIII Special Rapporteur Group on New Forms of Image Communication have been working together during 1987 and 1988 in a Joint Photographic Experts Group (JPEG) for the purpose of developing an international standard for the compression and decompression of natural color images. The technique selected is required to allow for both progressive and sequential image buildup during decompression. Decompression is to be feasible in real time in the ISDN environment (64 Kbits compressed data per second). The final standard is expected to produce a recognizable image at under 0.25 bits/pixel, an excellent image around 0.75 bits/pixel, and an image visually indistinguishable from the original around 3 bits/pixel for original images of 16 bits/pixel. Exact (lossless) coding is also required. An adaptive discrete cosine transform (DCT) technique was selected in January 1988 for further refinement and enhancement. The definition of the inverse DCT, how to improve the low-bit-rate image quality, the choice of entropy coding technique, and the method of achieving graceful progression are being studied as part of the refinement and enhancement process before final selection. A draft standard is expected to be available during 1989. The status of the refinement process will be reviewed.

Progressive coding1 for image data is the focus of current standardization activity within the CCITT and ISO organizations. Recently, a Joint Bi-level Image experts Group (JBIG) has been formed by these two organizations to study the question of progressive coding for bi-level image data. The point of departure for this work is the PCS algorithm for binary images described in ISO/CCITT standards meetings by Endoh and Yamazaki2, in which they added the use of arithmetic coding3 to previous proposals. Our paper summarizes some enhancements that can be made to this PCS scheme by improving the application of adaptive binary arithmetic coding (ABAC) technology. In particular, the "Q-coder" form of ABAC technology is used4 which was recently featured in the November 1988 issue of the IBM Journal of Reseaich and Development5.

This paper will briefly describe two well-known data compression algorithms Lempel Ziv (LZ) and Comte Consultatif International Telephonique et Telegraphique (CCITT G3 and G4) and will show the results of compressing numerous bit/pixel images from a number of different image categories using these two techniques. It will be shown that images compressed with the two techniques yield compression ratios which fall into two mutually exclusive regions which depend upon the entropy of the image. Moreover, the method giving the best compression ratio can be determined from a calculation based on the zero order statistics of the image. Thus a simple calculation can determine which of these two techniques will give the better compression ratio on a given image.

Realtime transmission of packetized video is a relatively new but important area of research since tomorrow's networks will likely be based on a common fast packet-switching technology, replacing today's heterogeneous networks which are mixtures of circuit-switched and packet-switched networks dedicated for synchronous and asynchronous applications respectively. Unlike fixed bandwidth transmission channels, packet-switched networks can accommodate the bursty and highly variable rate of compressed motion video yet keep a constant level of image quality. Potential applications include motion video databases, video teleconferencing, and broadcasting over private networks. Unfortunately, traditional video-compression techniques which are applicable to synchronous, fixed bandwidth channels--cannot be applied in this new environment without substantial revisions that take into account the statistical nature of packet data transmission. Independent of their topology, fast packet-switched networks present unique challenges to the coding of motion video; one cannot simply divorce the coding of the source from the characteristics of the channel. The potential for loss of data integrity due to packet losses and transmission errors must be taken into account. The problem of re-synchronization must also be addressed since packet-switched networks are asynchronous by nature. In this paper, we evaluate image coding techniques for motion video in view of the constraints imposed by the packetized medium.

Progressive image transmission provides rapid reception of an image approximation, which is thereafter refined without incurring excess coding overhead. Nonhomogeneous techniques improve progressive reception by concentrating early transmission efforts on the most informative image areas, again without incurring excess coding overhead. A nonhomogeneous progressive coding method is described for interframe or intersample (nonconsecutive frame) coding and transmission. Working upon the simple difference of successive frames/samples, the coding method first transmits the most important differences followed by differences of lesser importance. The amount of coded data and the transmission time for each sample difference is proportional to the degree of image change. The nonhomogeneous, progressive, intersample coding method does not provide motion compensation, but does have the advantage of simplicity and low computational requirements. As all progressive techniques, the coding method is particularly suited for use with low bandwidth communication channels.

High speed image compression algorithms that achieve visually lossless quality at low bit-rates are essential elements of many digital imaging systems. In examples such as remote sensing, there is often the additional requirement that the compression hardware be compact and consume minimal power. To meet these requirements a new adaptive differential pulse code modulation (ADPCM) algorithm was developed that significantly reduces edge errors by including quantizers that adapt to the local bias of the differential signal. In addition, to reduce the average bit-rate in certain applications a variable rate version of the algorithm called run adaptive differential coding (RADC) was developed that combines run-length and predictive coding and a variable number of levels in each quantizer to produce bit-rates comparable with adaptive discrete cosine transform (ADCT) at a visually lossless level of image quality. It will also be shown that this algorithm is relatively insensitive to fixed-pattern sensor noise and errors in sensor correction, making it possible to perform pixel correction on the decompressed image.

Real-time color image compression is always needed. This paper presents a practical error-tolerant compression algorithm for recording color pictures digitally on a tape and provides a real-time architecture such that the processing of picture compression is implemented in a single VLSI chip. The algorithm is based on the principle of block truncation coding (BTC). The picture is represented in Y-I-Q color space and each plane is divided into small blocks such that a reconstructed picture still contains the quality appropriate for 3R prints and keep compression efficiency. Any single channel error is restricted into a very small block of the picture and this feature of error-tolerance is important for the application of picture recording. The real-time architectures for three signal channels are working in parallel and each channel has a pipelinal architecture. This architecture also needs two 4-line input buffers with 24 bits in depth and two 96-bit output buffers. The whole architecture for both compression and decompression can be implemented with a single VLSI chip and be executed in real time. This approach provides two unique properties: error-tolerance and real-time execution, with which most other image compression algorithms have problems.

The need for a compression system to significantly reduce the size of high resolution gray-scale and full-color image (picture) files, which can range from 100 Kbytes to over 3 Mbytes, stimulated an extensive research program that resulted in the development of new image compression systems. These new compression systems, based on well established and suitable compression schemes which run on Graphic Signal Processor (GSP), can reduce image file sizes by a factor of more that 8:1 with only minor detectable image degradation. Image compression of such kind is of interest because GSP is now an integral part of display control. The need for separate and expensive DSP for Discrete Cosine Transform computation is eliminated. This paper will discuss the importance of this compression scheme in the GSP world in terms of cost, speed and performance.

This paper presents the results of the research that was conducted by Telephoto Communications into the effects of implementing the proposed ISO International Still-Frame Image Compression Standards. The objective of the research being to provide a commercial product which will comply with these standards. After providing a brief system description of the ALICE Type-120 and the "ISO compatible" product, the paper then reviews the modifications to the ALICE Type-120 algorithm to make it compatible with the proposed ISO standards and the performance impacts of these modifications. The primary elements of the ISO compatible system include: 1) Fast Discrete Cosine Transform (FDCT) 2) Differential Pulse Code Modulation (DPCM) of DC coefficients 3) Visual Quantization 4) Huffman encoding DPCM is a previous pel (pixel) predictor algorithm for coding DC coefficients. Visual quantization is a non-uniform quantization matrix based on the spectral position of each AC coefficient and takes advantage of the reduced sensitivity to high frequencies of the human visual system. Modifications to the ALICE Type-120 included the addition of the visual quantization and DPCM of the DC. The addition of these elements resulted in significantly improved compression ratios while maintaining a subjectively constant image quality. The research was conducted with the standard test images used in ISO evaluations and with special images used in the development of the ALICE Type-120 Image Compression System.

This paper discusses a hybrid coding method for low bit-rate video transmission over the 2B channels (128 kb/s) of ISDN networks. In this method the coding frame is divided into two dimensional m-by-m blocks where each block is further divided into smaller sub-blocks. Block classification and blockmatching motion estimation are based on the larger block size, while the transformation is performed on a sub-block basis. The transform coefficients are coded in such a way that the inter sub-block correlation can be efficiently exploited. As a result, for every block the sub-block coefficients are scanned by grouping all the first coefficients, followed by all the second coefficients and so on. The manner in which the sub-blocks are scanned is called inter sub-block scanning. Three types of sub-block scanning are considered: zig-zag, horizontal and vertical scanning. The inter sub-block scanning is made adaptive by deciding which of the above scanning types is the most efficient. Consequently, this requires additional overhead to be sent to the receiver at the beginning of each block. In our simulation the main block size of 16 x 16 and 32 x 32 is considered. In order to find the best sub-block size, four different sizes such as 2 x 2, 4 x 4, 8 x 8 and 16 x 16 are evaluated. The results using a few video sequences as the input show that a good quality video can be obtained at the transmission rates of 48 kb/s, 64 kb/s and 112 kb/s.

Full color printer by sublimation dye thermal transfer system was developed for proof printing. Proof printing contains various kinds of image such as characters, diagrams, pictures, etc. and it requires many contrast steps and high resolution. Especially, gradation technique is required for fine adjustment to obtain printed matter as close as original. We believed that this gradation technique is important. We realized the gradation of 256 contrast steps with constant density difference between any two adjacent steps by controlling heating time of thermal head. High quality proof is realized by our new invention for gradation technique where one contrast step is finely adjusted by 256 sub-steps. The major specifications of our printer include resolution of 11.8dots/mm, printing speed of 10 msec/line, and printing size of 302.8x439 mm.

In this paper we address the problem of displaying continuous tone photographic colour images on CRT monitors on which only a limited number of colours can be displayed simultaneously. An algorithm is presented which generates a palette of a limited number of colours, and a method is given for the actual display of a full colour image using such a palette and its associated tables.

Two color alteration algorithms for colorappearance simulation in design systems were developed. One algorithm is based on a surface reflection model and is suitable for plastic-like objects. Another algorithm is based on an absorption model and is suitable for transparent objects or textile-like objects. Applicability was evaluated and experiments were successful. High-speed algorithm and achromatic object coloring were also studied.

The implementation of 2-dimensional (2-D) linear transformations (e.g., Fourier and cosine transforms) is a computationally intensive task. In progressive image processing, it is desirable to use algorithms that produce outputs of 2-D transformations in successive steps throughout the computations. We propose an example of these algorithms for successive inversion of discrete cosine transformed images. The approach is suitable for interactive applications of progressive image presentation.

We discuss halftoning techniques appropriate for matrix-addressed electronic displays, such as thin-film-transistor (TFT) liquid crystal displays, which use color mosaics to implement multi-color operation via space-division-multiplexing. The error propagation method of halftoning is briefly reviewed and extended into the present context. Temporally dynamic techniques are described for rapidly refreshed displays, as is their implementation in digital hardware. Artifact suppression and the use of auxiliary passive light-optical image enhancement techniques are also addressed.

In the little known area of photo based visual systems (PBVS) for flight/sensor simulations, many traditional methods of Digital Image Processing are being actively used and extended and, new non-traditional Image processing methods are being developed. This paper describes the roles of both the traditional and new image processing methods, in PBVS, using the LTV system as an example.

Three alternative approaches to terrain rendering are presented in historical order. Polyhedral computer graphics, image warping, and image ray tracing are compared in terms of their relative strengths and weaknesses. Image ray tracing is found to be the most robust technique for most terrain rendering applications.

Recent advances in high speed film digitizing have made possible the extraction of digital image data from high resolution film at rates equal to those of the fastest digital storage media. The major features of this technology are presented and performance characteristics are given.

Preparing data bases for highly-realistic perspective scene generation has several difficult aspects, including mosaicking images and representing vertical features. A perspective scene is a digital image generated to simulate viewing the ground from a selected observation point, looking in a selected direction. Each perspective scene is generated from a data base containing digital image data and grid elevation data. In many cases, several digital images must be combined or mosaicked to produce the image data needed for generation of one or more perspective scenes. The image mosaicking requirements to prepare data bases for perspective scene generation are described. Alternative mosaicking approaches are then reviewed, identifying the more attractive methods. A semi-automatic mosaicking process using these methods is outlined. Digital images have been successfully mosaicked using this process, and experimental results are shown. The digital image data is normally vertical or orthorectified, and thus cannot directly represent the appearance of vertical and overhanging sides of ground features. The requirements are described for representing the sides of features to adequately support realistic perspective scene generation. Alternative feature representation approaches are then reviewed, selecting the more attractive methods. Planar feature surfaces can be extracted, including the surface location in ground coordinates and image data showing the appearance of the surface. A feature data format and corresponding perspective scene generation algorithms are outlined, and experimental results are shown.

Image viewing applications often require features not found in display systems designed for general graphics applications. These differences are mostly due to the fact that in imaging applications the data already exists in the form of pixels and must be moved into a frame buffer, whereas in graphics applications pixel data is generated locally and I/O consists of high level commands or display lists. When performing functions such as zoom, split screens, and virtual roam (a technique for viewing images larger than the local frame buffer), an image display system requires hardware implementations to perform these functions for maximum performance. This special hardware involves the data path between the frame buffer and the video output. For graphics applications the display list can be modified, and with a reasonably fast graphics processor no special video data path hardware is required. This paper describes a hardware windowing scheme used in an image display processor which allows multiple images to be displayed simultaneously, each with differing zoom factors, LUT settings and pixel size. The architecture is powerful enough to allow simultaneous display of up to sixteen virtual roam windows. The display processor is supported by a high bandwidth bus and memory system allowing more I/O intensive applications such as film looping to be performed.

A traditional problem in real time perspective image generation applied to data bases covering large real geographic areas has been the compromise between the desire for realism and system cost. We present some innovative solutions to this problem using (i) constant time ray tracing software, (ii) the transformation of real image data pixels into voxels, and (iii) parallel processing.

This paper addresses to provide some of the new image processing techniques for color hardcopy. The first requirement for high quality color imaging, is how to reproduce a continuous tone. In the actual printing systems, we often encounter the false contours caused by a lack of gray levels or a jump of dot density. Here, we present a tone smoothing method using an improved pseudo-noise modulation as a countermeasure for such a printer with discontinuities in gray scale reproduction. The second requirement is to get the correct color rendition. We discuss two models on color correction and show the practical applications for full color printing. As the third requirement, a function to adjust color-tone as you like, will be needed for the future hardcopies and we shortly discuss the color adjusting problem based on human color perception. Lastly, we report our latest works on the new color coding technique, which can reconstruct a full color image from a single color image by utilizing the strong correlation characteristics among RGB signals.

Resampling an image will be described as a low-pass filtering operation followed by sampling to a new coordinate system. To determine the interpolation function that gives the most visually appealing images, a comparison of common kernels is made. Linear, cubic, and windowed sinc functions are compared in terms of frequency response and with prints of images resized using separable extensions of these functions. While the windowed sinc gives the best approximation to an ideal low-pass filter, using this kernel results in objectionable ringing and jaggedness around edges in the image. Cubic interpolation is shown to provide the best compromise between image sharpness and these edge artifacts. For image rotation, and resizing by an arbitrary factor, the filter coefficients (samples of the interpolation function) need to be computed for each pixel of the new image. Alternatively, significant computation can be saved by dividing the distance between pixels of the original image into a number of intervals and precomputing a set of coefficients for each interval. Each new pixel is then computed by finding the interval in which it falls and using the corresponding set of coefficients. An analysis of the errors introduced with this resampling method is presented. This analysis shows the number of intervals required to produce high-quality resampled images.

A review of the halftone processing techniques used in the Xerox 7650 Pro Imager document scanner is presented. The halftone processing techniques which are necessary for high quality, continuous tone document scanners to function in lithographic, xerographic and similar graphic arts environments are described. These processing techniques include image histogram manipulation, electronic screening, halftone removal and image enhancement. A method of automatic contrast adjustment which compensates for low contrast continuous tone or colored background images is shown. The need for programmable halftone generation to adapt to a wide variety of printer reproduction characteristics is discussed. A programmable halftone removal technique, and its application to halftoned pictorials, is addressed from a practical perspective. This discussion focuses particularly on the problem of the moire encountered when applying halftone patterns to scanned images which contain halftone pictorials themselves. Utilizing several example images, it is shown that by combining the described halftone image processing techniques available in the 7650 scanner, high quality, hardcopy images can be obtained on a variety of high contrast printers from a variety of continuous tone and halftone input documents.

On this paper, we set out a system which deals with the extraction of speed and position vectors of a target in an image sequence with standard technological elements within a minimum time. The application field, we want to deal with is the one of robotics and the automatic tracking of vehicules. The strategy which is used to build such a system requires three complementary techniques : optics, electronics and data processing of which respective functions are filtering, the operation of difference and the calculation of dynamic parameters. The speed of optical treatments combined with electronics leads to a real time isotrope edge detector, it's space cut-off frequency being adjustable. Based on the unsharp masking technique this extractor carries out the difference between a sharp image and the same image which has been previously defocussed. Therefore this modified image is quantized according to statistical criteria. This process optimizes the number of quantization levels useful for a given S.N.R. in order to reduce the size of the digital image. It means to reduce further calculation time. At last a fast template matching algorithm is carried out from partial correlation functions. These functions are calculated for each level of quantization and modified by a relative weighting. Lastly, a simple numerical calculation assesses the speed and position of the target. The performances of this process allow us to reach calculation times below 200 ms on a standard 3 MIPS processor when dealing with most of the tracking problems and in presence of "noise" as well. For instance, this device can either measure the automatic tracking of an all roads vehicule in a field, or the automatic hold of an object moving on an assembly line thanks to a robot.

As far as target tracking is concerned in the robotics field, the picture processing phase consists in finding out the location of the object in the scene within a minimum of time. Therefore, we require quite a new approach of "restrictive correlation" which combines optical function (real time edges extraction) and data processing functions (multilevel correlation). This latter as to show the best matching place between a model of the object and the search area. It consists of a multilevel thresholding followed by a model image mapping at each level. We find out the localization of the object by a weighting addition of each obtained result. The time required to obtain such a result is directly linked to the number of selected points at the thresholding stage.Therefore, we develop an analytical method to count the treated points according to the threshold levels. The grey levels of the picture's points are taken as a realization of a random process of which we measure the statistical characteristics (mean, standard deviation). If we refer to the theory of signal processing, this enables us to determine, by means of calculation, the number of points over a given threshold within an image for a kind of scene. These calculations are carried out for various levels. Afterward, their results are compared with the figures experimentally measured. On this way, we valid a relation which links the execution time of a correlation to its parameters. Consequently, this evaluation gives a quantitative criterion for the values which point out the limits with regard to the choice of thresholds ; the time available for the correlation being previously defined according to the amplitude of the search area and the maximal speed authorized for the target.

In this paper we examine the use of projection techniques for image restoration. All the methods are shown to be specific examples of the method of projection onto convex sets (POCS). The projection techniques that commonly project only onto linear subspaces are generalized to include projections onto nonlinear subspaces. A new technique based on POCS is described and shown to yield better restoration results than the other projection techniques.

A technique for the semi-automatic detection of cell and nuclear contours in cytological specimens is presented. Gray level images are viewed as stacked binary bit plane images with each bit plane image representing the effect of thresholding the gray level image with multiple gray level windows. A binary morphological edge detector is employed to identify edges on selected bit planes. A simple contour tracking algorithm is used to identify the most likely cell and nuclear contours which are presented to the user for verification.

A hybrid correspondence algorithm is presented to improve a previously developed framework for estimating motion of a moving object from its range image frames. Range images are represented by relational graphs in which a node indicates a view-invariant patch on the object surface and an edge the adjacency of two such patches. To a node, sign of Gaussian curvature and to an edge, distance between centers of mass are assigned as attributes. To estimate motion between consecutive range image frames, a correspondence algorithm is required to detect the largest common subgraph in the two representing relational graphs, i.e. to detect the surface part visible in both range image frames. The developed hybrid algorithm overcomes matching ambiguities for those graphs having a star structure. It combines the advantages of the previously developed correspondence algorithms by considering both the adjacency between a patch and its neighbors and the adjacency among neighbors themselves.

This paper discusses a system that integrates image processing and neural networks in a single high speed computational system. It then stresses the importance of performing the image preprocessing, feature extraction and neural network based training and recognition under the same platform. The efficiency provided by such a system is of vital importance for data and compute intensive image pattern recognition problems.

This paper is concerned with a systematic exposition of the usefulness of two-dimensional (2-D) discrete Gaussian Markov random field (GMRF) models for image processing and analysis applications. Specifically, we discuss the following topics; notion of Markovianity on a plane, statistical inference in GMRF models; and their applications in several image related problems such as, image synthesis, texture classification, segmentation and image restoration.

A neural network architecture is presented in which the strengths of interconnection pathways in one part of the network are modulated by activities in another part of the network. The network is arranged in a series of stages forming a hierarchy. The lower level stages produce output codes that are abstractions of the desired categorization task. Multiple levels of abstraction combine to provide categorization and generalization capabilities that appear to exceed those of standard backpropagation. Empirical results on two-dimensional continuous-valued (analog) simulated feature data are presented.

This paper describes a rigorous approach to radar vision. The two foundations of the approach are a theory for handling the structure and uncertainties found in radar vision and a testbed implementation of portions of that theory. The theory is novel and powerful in that it explicitly accounts for uncertainties in object modeling and electromagnetic phenomenology modeling. A key concept is that the vision problem itself is modeled as a highly structured, probabilistic, joint parameter estimation problem involving many observables and unknowns. The parameters are both discrete, such as existence or type, and continuous, such as location or shape. The problem structure involves widespread conditional decoupling of effects and allows for efficient, near-optimal algorithms for inferring world parameters from observations.

Automatic Target Recognizers (ATRs) are required to improve the performance of future missile systems to enhance performance against evolving threats and to increase cost effectiveness of future "smart/brilliant" missiles, aircraft and submunitions. This paper will address the present status of ATRs and indicate areas which require new techniques to improve system performance.

The design and application of a custom IC (integrated circuit), developed to enhance an existing real-time image processing system, is described. The system is used to automatically monitor road traffic at a complex road junction, where data is acquired using a black and white video camera. The new custom IC forms part of a preprocessing circuit to remove binary noise from the image, and is particularly effective when used in conjunction with a technique for background updating [1], also described.

In general it can be said that the textile industry endeavours to render a bunch of fibers chaotically distributed in space into an ordered spatial distribution. This fact is independent of the nature the fibers, i.e., the aim of getting into higher order states in the spatial distribution of the fibers dictates different industrial processes depending on whether the fibers are wool, cotton or man made but the all effect is centred on obtaining at every step of any of the processes a more ordered state regarding the spatial distribution of the fibers. Thinking about the textile processes as a method of getting order out of chaos, the concept of entropy appears as the most appropriate judging parameter on the effectiveness of a step in the chain of an industrial process to produce a regular textile. In fact, entropy is the hidden parameter not only for the textile industry but also for the non woven and paper industrial processes. It happens that in these industries the state of order is linked with the spatial distribution of fibers and to obtain an image of a spatial distribution is an easy matter. To compute the image entropy from the grey level distribution requires only the use of the Shannon formula. In this paper to illustrate the usefulness of employing the entropy of an image concept to textiles the evolution of the entropy of wool slivers along the combing process is matched against the state of parallelization of the fibbers along the seven steps as measured by the existing method. The advantages of the entropy method over the previous method based on diffraction is also demonstrated.

The raw material for the paper industry is wood. To have an exact account of the stock of piled sawn tree trunks every truck load entering the plant's stockyard must be measured as to the amount of wood being brought in. Weighting down the trucks has its own problems, mainly, due to the high capacity of the tree trunks to absorb water. This problem is further enhanced when calculations must be made to arrive at the mass of sawn tree trunks which must go into the process of producing a certain quantity of paper pulp. The method presented here is based on two fixed cameras which take the image of the truck load. One takes a view of the trunks in order to get information on the average length of the tree trunks. The other obtains a side view which is digitised and by just discriminating against a grey level the area covered by the tree trunk cross section is measured. A simple arithmetic operation gives the volume of wood in the trunk. The same computer, a PC, will register the trucks particulars is almost independent of weather the wood is wet or dry and it serves trucks of any size.

A system has been developed by which colour changes may be detected and measured over the whole surface of a painting. The equipment, which comprises a solid state camera, frame store and computer has been used to acquire, correct and compare colour separation images of paintings.

The Bulk Image Transformation Engine (BITE), developed at International Imaging Systems, has been designed specifically to address the unique problems encountered in processing large, multi-band, image data-sets. The key elements of the hardware and software architecture that combine to provide an efficient, yet easy to program, solution to these problems will be discussed. A generalized image warping algorithm developed for the BITE illustrates the performance of this architecture on a real-world problem.

A method for determining percent cloud cover using electronic imaging and physical optics of light scattering is described. This method involves taking two consecutive images using different color separation photographic filters, specifically one with a pass band in the blue region and one with a pass band in the red region. The "red" image is then divided by the "blue" image, pixel by pixel, and the percentage of pixels in the resultant image with grey levels equal to one is determined. This is the percent cloud cover. Estimates of percent cloud cover using this method were compared to visual estimates using the northern portion of 100 Central Florida skies. A strong correlation (r = 0.94) was found between the two estimates. Follow-up work on the prototype system is described.

A range of image processing hardware and software is used as a research tool and provides useful data for a number of Defence projects. Often these projects require the development of original ways of using and calibrating state-of-the-art equipment to record images and process them to aid interpretation and provide the required data. Initially, image acquisition and digital image processing techniques were used at Materials Research Laboratory (MRL) to observe phenomena that occurred during the high speed transient events associated with energetic materials such as explosives, pyrotechnics and propellants. However, with the introduction of automated image processing techniques, more diversified applications such as particle sizing and fragmentation sizing studies have become commonplace. This paper discusses the applicability of image enhancement, processing and measurement to projects such as the study of shaped charge warheads. Measurements relating to shaped charge warhead formation and penetration have provided a data base for assessing the accuracy of 2-D computer model codes. Digital image processing techniques have also been valuable in assessing the ability of multispectral screening smokes to camouflage Defence vehicles and personnel. Other applications incorporate image analysis techniques used in conjunction with flash radiography, high speed video, and photo instrumentation imaging devices. Discussion includes experimental configuration and calibration of these devices so that geometric and grey level integrity of subjects is controlled and optimised prior to, and during, digital image processing.