The determination of an objective quality scale for image coding that corresponds to the subjective evaluation of quality is a long standing problem that has eluded researchers for many years because of its complexity. As a consequence, SNR, which is recognized as an inappropriate method for evaluating image coding, has been widely used for convenience. Recently, several of the authors have developed a new approach to the quantitative determination of image quality, defining a picture quality scale (PQS) that correlates well with subjective ratings, for mean opinion scores (MOS) in the range of 2 to 4. In this work, we extend and specialize these results for the case where the quality requirements become very high and where the distortion should be barely perceptible. We consider the detailed spatial distortion maps, which are the local contributions for distortion factors that make up PQS, as spatial indicators of perceptible distortions. We determine the applicability of these distortion maps to the type of distortions that remain perceptible at high quality. We propose some modifications to the PQS metric that would apply at high quality. The experimental results are established for images processed using the JPEG coding standard.

In the perceptually transparent coding of images, we use representation and quantization strategies that exploit properties of human perception to obtain an approximate digital image indistinguishable from the original. This image is then encoded in an error free manner. The resulting coders have better performance than error free coding for a comparable quality. Further, by considering changes to images that do not produce perceptible distortion, we identify image characteristics onerous for the encoder, but perceptually unimportant. Once such characteristic is the typical noise level, often imperceptible, encountered in still images. Thus, we consider adaptive noise removal to improve coder performance, without perceptible degradation of quality. In this paper, several elements contribute to coding efficiency while preserving image quality: adaptive noise removal, additive decomposition of the image with a high activity remainder, coarse quantization of the remainder, progressive representation of the remainder, using bilinear or directional interpolation methods, and efficient encoding of the sparse remainder. The overall coding performance improvement due to noise removal and the use of a progressive code is about 18%, as compared to our previous results for perceptually transparent coders. The compression ratio for a set of nine test images is 3.72 for no perceptible loss of quality.

In this paper, a DCT based coding technique for adaptive coding of images according to the level of visual activity is presented. Adaptation is based on adaptive quantization and adaptive bit selection. In the proposed system, we initially partition the image into a large number of sub-blocks of 4 X 4 pixels. A novel image analysis may then be performed prior to the coding in order to decide what is the most significant information to encode. Classification according to the activity level within the blocks is based on the local statistics, and is used for adaptive bit selection, whereas optimum quantifiers having Gaussian density are used to achieve adaptive quantization. Satisfactory performance is demonstrated in terms of direct comparison of the original and the reconstructed images.

This paper describes a general approach to interframe video coding using a method of local feature-based motion compensation. Used in conjunction with a DPCM/DCT interframe coding technique, such as the CCITT H.261 algorithm, this technique is computationally viable at real-time rates, employing a simple process of evaluating the nature and content of features and their subsequent displacement. As a secondary means of interframe coding, it has proved possible to introduce mode-value blocking and quantization to speed the motion estimation process and results indicate that as part of a hybrid encoding algorithm, this approach is quite acceptable as no long-term errors are propagated. In addition to a description of feature analysis as a combinational coding method, it is shown that the process of feature extraction and evaluation provides, in its own right, scope to form the basis of a complete, efficient codec of low complexity.

Future land remote sensing satellite systems will likely be constrained in terms of downlink communication bandwidth. To alleviate this limitation the data must be compressed. In this article we present a robust and implementable compression algorithm for multispectral imagery with a selectable quality level within the near-lossless to visually lossy range. The three-dimensional terrain-adaptive transform-based algorithm involves a one-dimensional Karhunen-Loeve transform (KLT) followed by two-dimensional discrete cosine transform (DCT). The images are spectrally decorrelated via the KLT to produce the eigen images. The resulting spectrally-decorrelated eigen images are then compressed using the JPEG algorithm. The key feature of this approach is that it incorporates the best methods available to fully exploit the spectral and spatial correlation in the data. The novelty of this technique lies in its unique capability to adaptively vary the characteristics of the spectral decorrelation transformation based upon variations in the local terrain. The spectral and spatial modularity of the algorithm architecture allows the JPEG to be replaced by a totally different coder (e.g., DPCM). However, the significant practical advantage of this approach is that it is leveraged on the standard and highly developed JPEG compression technology. The algorithm is conveniently parameterized to accommodate reconstructed image fidelities ranging from near- lossless at about 5:1 compression ratio (CR) to visually lossy beginning at around 40:1 CR.

An algorithm class called CaTH (centering and tail handling) is described that is based on predictive coding followed by adaptive binary arithmetic coding. CaTH treats the prediction errors close to zero (i.e., near the center of the error distribution) in a more precise manner than the errors of the `tails' (i.e., errors far from zero). The context model uses error buckets (quantized ranges) of prediction errors. The probability model for the prediction errors uses a histogram for the center. A variety of ways to binarize the tails are studied. The results on the suite of JPEG test images are very encouraging.

Encoding of images with requirements of high visual quality and post-processing capability requires new approaches to the coding problem. These requirements prohibit the introduction of noticeable degradation and structured artifacts by the coding process. A variable block size coding approach using a spatial structure criteria for segmentation has been found to yield high compression ratios, particularly for images with detail concentrated in limited regions. However, block based techniques, such as transforms, vector quantization (VQ) or a combination of transforms and VQ, by their very nature tend to introduce unacceptable blocking artifacts. In this paper, a modified variable block size approach tailored to yield high quality and no structured artifacts is developed. The technique is used in a two-source coding scheme and in a multi-source scheme i.e., coding of the Laplacian pyramid decomposition. Results of simulations on standard 512 X 512 monochrome images are obtained and compared with those obtained by the baseline JPEG standard and current literature. Results of simulations on 1024 X 1024 digital x rays are also obtained. The proposed two-source scheme is found to yield good compression ratios, high quality and impressive post-processing capability.

This paper proposes a new ADPCM method for image coding called directional ADPCM which can remove more redundancy from the image signals than the conventional ADPCM. The conventional ADPCM calculates the two-dimensional prediction coefficients by using the correlation functions and solving the Yule-Walker equation. Actually, the quantities of correlation functions are replaced by the sample averages. Therefore, this solution will not be optimum. Our directional ADPCM utilizes the directional filters to obtain the energy distribution in four directions and then determines the four directional prediction coefficients. All the directional filters are designed by using the singular value decomposition (SVD) method and the two-dimensional Hilbert transform technique. In the experiments, we illustrate that the M.S.E. for the directional ADPCM is less than that of the conventional ADPCM.

Although much work has been done towards developing lossless algorithms for compressing image data, most techniques reported have been for two-tone or gray scale images. It is generally accepted that a color image can be easily encoded by using a gray scale compression technique on each of the three (say, RGB) color planes. Such an approach however fails to take into account the substantial correlations which are present between color planes. Although several lossy compression schemes that exploit such correlations have been reported in the literature, we are not aware of any such techniques for lossless compression. Because of the difference in goals, the best ways of exploiting redundancies for lossy and lossless compression can be, and usually are, very different. In this paper we propose and investigate a few lossless compression schemes for color images. Both prediction schemes and error modeling schemes are presented that exploit inter-frame correlations. Implementation results on a test set of images yield significant improvements.

The paper reports on a video sequence coding method taking advantage of the generic video- communication layout: some moving objects on a still background. The algorithm operates on groups of frames in which the whole digital video sequence is divided, that implies the synchronization requirements' satisfaction and an acceptable level of compatibility with standard video coding (H.261, MPEG, etc.). An analysis of the spatial-temporal continuum, represented by each group of frames, is performed, in order to detect a tridimensional segmentation that identifies the moving objects by means of spatial regions. These regions can spread, as a sort of `pipes,' through the whole group of frames in the temporal direction. Various pipes' construction and coding strategies, including techniques based on object recognition and coding, are allowed. In this work a pipes' identification method based on fixed size moving blocks and their coding by means of a 3D-DCT transform is reported. The above method allows adjacent starting pipes to part themselves, leaving uncoded stripes at their boundaries. The proposed method does not imply the stripes coding, while it minimizes their number and the amount of the artifacts generated by their presentation. As a final topic, the paper reports some considerations on the coding efficiency related to the quality of the reconstructed sequences and on the compatibility characteristics.

Enhancement of both multispectral and true-color images is often performed by histogram modification, usually by separately adjusting the color components within a selected color coordinate system. When preservation of certain perceptual qualities is important, a perceptually based coordinate system is employed. However, independent modification of the color components seldom results in full use of the RGB display gamut unless some color values are clipped at the RGB boundaries. Preserving perceptual attributes is sometimes less important than obtaining greater displayable color contrast. This is especially true for color composites derived from multispectral images, such as those obtained by remote sensing. `Histogram explosion' is a new, multivariate enhancement method able to exploit nearly the full RGB extent without causing clipping. It functions by deriving and adjusting a set of one- dimensional histograms which lie along rays emanating from a common point in RGB space. While not generally based upon a perceptual model, histogram explosion is capable of preserving original hue values when parameters are chosen properly. Wide flexibility in the algorithm's parameter choices allows much freedom in tailoring the enhancement process. A description of these parameter dependencies and an analysis of the computational complexity are presented.

An information divergence, such as Shannon mutual information, measures the `distance' between two probability density functions (or images). A wide class of such measures, called (alpha) -divergences, with desirable properties such as convexity over all space, has been defined by Amari. Renyi's information D_(alpha ) is an (alpha) -divergence. Because of its convexity property, minimization of D_(alpha ) is easily attained. Minimization accomplishes minimum distance (maximum resemblance) between an unknown image and a known, reference image. Such a biasing effect permits complex images, such as occur in ISAR imaging, to be well reconstructed. There, the bias image may be constructed as a smooth version of the linear. Fourier reconstruction of the data. Examples on simulated complex image data, with and without noise, indicate that the Renyi reconstruction approach permits super-resolution in low-noise cases, and higher fidelity over ordinary, linear reconstructions in higher-noise cases.

In this research, we propose a new scheme to integrate the two important depth information sources, i.e., shading and stereo, in a single framework. By using the ratio of two photometric stereo images, we first derive a new constraint on the local surface structure so that no albedo information is needed for shape reconstruction. Then, by employing the perspectively projected SFS formulation with a generalized parametric surface model, we establish a unified geometrical framework for the integration of shading and stereo. The depth map is directly recovered by minimizing a cost functional which consists of a weighted sum of shading and stereo error terms. Simulation results are given to show the performance of our new robust algorithm.

Many space based optical systems, especially in the infrared, have the constraint that the focal plane cannot be adequately sampled due to IR detector constraints. Adequately sampled refers to sampling at twice the optical MTF cutoff frequency. Generally the focal plane contains multiple columns of detectors with each detector element being large with respect to the system PSF. This necessitates a system which is scanned either by moving a scan mirror or physically moving the spacecraft or a combination of both. Discussed is a method which allows rapid construction, from multiple column scan data, of an image sampled on a square grid with minimal modification of frequency content up to the spatial Nyquist frequency of the image grid. The method relies on a modified implementation of the 2-D Whittaker-Shannon sampling theorem to minimize aliasing and `spectral leakage.' The algorithm and its implementation for the Spirit III radiometer of the Mid-Course Space Experiment (MSX) are discussed along with a rapid implementation on a multiple processor desktop workstation.

Image restoration deals with images in which information has been obscured or partially lost. Practical problems are usually ill-conditioned. In this paper, we present an adaptive, iterative approach for these kinds of problems. The proposed approach is implemented in space domain, and it updates a restored image and an estimate of the regularization parameters simultaneously at each iteration. No prior knowledge about the noise variance is assumed. A space-variant operator, which works based on local information, determines the regularization parameters in order to ensure the regularization is `tight' in the smooth regions but `loose' in edge regions. This approach can be used for more general and spatially varying cases. Linear and non-linear constraints can be incorporated into the iterations. These constraints are motivated by the objective of accomplishing restorations with reduced artifacts such as rings and filtered noise artifacts, and preventing sharp edges while enhancing detailed structures in images simultaneously. The performance of the method is analyzed and both visual inspections and numerical results are presented. For example, SNR improvement of (Delta) SNR equals 8.48 dB has been achieved for the defocused 256 by 256 Cameraman image (7 by 7 2-D uniform blurring, BSNR equals 40 dB) and (Delta) SNR equals 3.72 dB has been obtained for the BSNR equals 20 dB image.

The recognition of objects that are imaged by viewing through a refractive interface has long been considered a difficult goal, due to spatial distortions that result from interfacial refraction. Although not insoluble via approximation, trans-interfacial stereophotogrammetry has not been addressed in the open literature. However, our previously-reported research has presented feasible, computationally efficient methods for restoring imagery obtained by monocular trans-interfacial viewing. The restoration results exhibit visually acceptable degradation due to information loss inherent in the restoration algorithm and (by design) a partial knowledge of the interfacial optical parameters. In particular, we have shown that our methods can facilitate low-distortion viewing of objects that are submerged at shallow optical depths (less than one caustic surface) in moderately clear water. In practice, our techniques require a partial knowledge of the sea topography, which we propose to obtain through time- domain reflectometry, or via stereophotogrammetry of the air-water interface. In this paper, we extend our previous work in monocular trans-interfacial imaging to address the problem of determining the range-to-target in trans-interfacial stereoscopic imagery. Note that our previous scenarios for interfacial viewing, which assumed bistatic sensing of interfacial topography, required at least three high-resolution cameras. Unfortunately, such techniques present infeasible bandwidth requirements for operationally-available transmission channels. Additionally, sensor design is complex, and performance can degrade significantly in the presence of slight optical misalignment. Thus, we restrict our discussion to the problem of stereo vision through the interface using only two cameras that exhibit moderate resolution. Analyses emphasize determination of the constraints upon sensor configuration and sea state under which such methods are applicable, as well as the prediction of spatial localization errors in the presence of typical sensor parameter errors and limiting cases.

The distribution of edge values for an image of a general scene often has a sharp peak with a long tail. This property which can be well described by a Lorentzian probability function has been used to develop an efficient non-linear image restoration algorithm for reducing the various artifacts that often arise in the restored images. The algorithm starts with a Wiener filter solution which is used to model the edge image by the Lorentzian function so that the likelihood of the image can be estimated. A non-linear correction term is then introduced which increases this image likelihood under the mean square error (MSE) criterion. This process ensures that the resulting image retains its sharpness while reducing the noise and ringing artifacts. An iterative procedure has been developed to implement this method. Computer simulated results show that the algorithm is robust in reducing artifacts and easily implemented. The algorithm also possesses a superresolution capability due to the highly nonlinear property of the correction term.

This paper is concerned with removing a typical noise artifact: the line patterns that arise from the use of parallel detector channels in an imager. There is an advantage in using a number of parallel scanned radiometers in order to reduce the frame time of an imager, but the response of these channels tends to be non-uniform and produces regular scan lines in the image. This effect may conceal important image details and reduce the signal to noise ratio in the image with the result that image restoration algorithms are not as effective as they could be. A number of efficient methods have been developed to remove these scan lines, either in real space or in the Fourier domain. These techniques can be applied to both infrared and millimeter wave systems and their effectiveness has been demonstrated by practical examples.

The conventional printing technologies use the standard four ink (yellow, magenta, cyan, and black) process. In some special cases, an unconventional ink set printing process can be selected. The unconventional inks set printing process uses a particular set of inks, selected according to the image to be reproduced. The inks are deposited on the substrate material in a certain order, through individual ink masks. Designing the unconventional printing process requires us to find the ink color set, named primary color palette, the sequence of ink mask printing, and the individual ink masks (color separation). In addition to the primary color palette, the secondary color palette is defined as the set of all the color combinations resulting by overlapping the ink masks after the complete printing process. A single primary color palette may conduct after printing process to a number of different secondary color palettes. This paper provides a color separation method which converts the input color components to the ink mask values that determine the amount of ink deposited on the substrate. The separation process depends essentially on the color calibration process that determines the colors of the secondary palette. The color calibration is accomplished using the printed samples. The calibration process conducts to a hierarchical structure of color combinations of the secondary color palette. The simulation of the printed colors is described, based on the generalized Neugebauer model for an arbitrary number of inks and an arbitrary set of inks. The color hierarchy established during the calibration process is used to derive the coefficients of the Neugebauer model.

The technique of mapping an array of gray levels to some arrangement of dots such that it renders the desired gray levels is called halftoning. In this research, we present a refinement of our previously proposed new digital halftoning algorithm to achieve this goal based on an approach called the recursive multiscale error diffusion. Our main assumption is that the resulting intensity from a raster of dots is in proportion to the number of dots on that raster. In analogy, the intensity of the corresponding region of the input image is simply the integral of the (normalized) gray level over the region. The two intensities should be matched as much as possible. Since the area of integration plays an important role to how successful the matching of the two intensities can be, and since the area of integration corresponds to different resolutions (therefore to different viewing distances), we address the problem of matching the intensities, as much as possible for every resolution. We propose a new quality criterion for the evaluation of halftoned images, called local intensity distribution, that stems from the same principle i.e., how close the average intensities of the input and output images match for different resolutions. Advantages of our method include very good performance, both in terms of visual quality and when measured by the proposed quality criterion, versatility, and ease of hardware implementation.

Histogram equalization is a well known tool for enhancing the contrast and brightness of grayscale images. Grayscale histogram equalization has been extended to color images with limited success. One common method is to equalize the illumination component, while leaving the saturation and hue components unchanged. This method doesn't improve the overall color saturation of the image. Another approach is to apply equalization techniques in the RGB color space. The difficulty in using the RGB color space is that it does not correspond to human interpretation of color. This paper describes a method for histogram equalization of the saturation component using the color difference (or C-Y) color space. Since equalization of the saturation component alone leads to color artifacts, attention is given to the relationship that exists between saturation and intensity.

Modern food production techniques operate at high speed and sometimes fill several containers simultaneously; individual containers never become available for inspection by conventional x- ray systems. There is a constant demand for improved methods for detecting foreign bodies, such as glass, plastic, wood, stone, animal remains, etc. These requirements lead to significant problems with existing inspection techniques, which are susceptible to noise and are unable to detect long thin contaminants reliably. Experimental results demonstrate these points. The paper proposes the use of two x-ray inspection systems, with orthogonal beams to overcome these difficulties.

Recently, calculation speed of computers has become faster and capacity of archival media has become larger. These situations enable image databases which efficiently archive, retrieve and utilize large amounts of image data. In order to retrieve required images efficiently from large amounts of image data, it is necessary to extract effective keywords from original images automatically. Many methods have been proposed for this purpose. However, satisfactory methods have not been found yet. In this paper, a new image data management method using vector expression of hue distribution of color images is proposed. A kind of normalized histogram of hue of a color image is regarded as a vector which represents image features. This unit vector of hue is utilized as a basis of the keywords of the image. The proposed image data management method enables flexible interface for various requests on retrieval. For example, `navigation in image data space based on similarity,' `retrieval of partial images or similarity retrieval using partial images,' `retrieval by semantic description,' such as `mountain view with clear sky and green forest' can be realized using the proposed method. The results of simulation experiments reveal the effectiveness of the proposed method.

The process by which images are prepared for complex vision systems is nontrivial. In this paper we describe the sequence of steps required to preprocess a grayscale picture for input to the Viewpoint Independent 3-D Extraction and Recognition of Objects (VITREO) system. VITREO is capable of accepting grayscale pictures for processing into line drawings for input to object recognition subsystems. These drawings are analyzed for edge and surface features to allow the extraction of component parts for subsequent recognition of the object containing them from stored component descriptions. This analysis demands that the line drawing extraction be robust and a combination of edge and line algorithms are employed. This process is described and examples of VITREO object extraction are shown.

The inspection and evaluation of surface mount assemblies and solder joints is a complex problem in computer vision which is as yet not satisfactorily solved. In this paper we outline an approach to a fast, low-cost inspection system, using a simple optical platform as opposed to more complex alternatives such as x-ray techniques or laser range finding. This paper describes the software modules required to perform feature extraction and shape analysis for inspection. In particular, we give detailed attention to the process of creating fully enclosed shape boundaries avoiding gaps in the outline of an object where the contrast is weak or noisy. A novel multi-stage generic linking and dropout correction algorithm is presented which preemptively overcomes problems difficult to rectify at later stages within the vision software. The algorithm is illustrated using examples from the surface-mount inspection system.

This paper describes the use of intelligent image processing as a machine tool operator safety technology. One or more color, linear array cameras are positioned to view the critical region(s) around a machine tool. The image data is processed to provide indicators of an operator danger condition via color content, shape content, and motion content. The data from these analyses is then sent to a threat evaluator. The purpose of the evaluator is to determine if a danger condition exists based on the analyses of color, shape, and motion, and on `knowledge' of the specific environment of the machine tool. The threat evaluator employs fuzzy logic as a means of dealing with uncertainty in the vision data.

Wavelet matrix bases can be interpreted as collections of windows that are complete and orthonormal. This paper describes a construction for windows with low sidelobes and rapid asymptotic decay, based on perfect reconstruction filters and wavelet matrices of infinite rank.

High quality lossy image compression schemes which provide high compression rates are an important tool in enabling practical picture archiving communication and storage (PACS) systems for medical radiology. The most widely used lossy schemes have been transform based, typically using the discrete cosine transform (DCT) as a first step. Newer transform methods based on the wavelet transform have demonstrated higher compression for a fixed quality level. A discussion of the basic method common to all transform based image compression schemes is presented. Evaluation criteria for lossy compression schemes are discussed, and the criteria are applied to both DCT (JPEG) and wavelet based methods, concluding that the wavelet based methods provide better quality and higher compression for most source imagery, radiological and otherwise.

Previous work has demonstrated the effectiveness of the expectation-maximization algorithm to restore noisy, degraded single-channel images and simultaneously identify its blur. In addition, a general framework for processing multi-channel images using single-channel techniques has also been developed. This paper combines and extends the two approaches so that simultaneous restoration and blur identification is possible for multi-channel images. However, care must be taken in estimating the blur and the cross-power spectra, which are complex quantities. With this point in mind, explicit equations for simultaneous identification and restoration of noisy, blurred multi-channel images are developed, where the images may have cross-channel degradations. Experimental results are shown which support this multi- channel approach, and are compared with multi-channel Wiener filter results. Independently restoring each channel is also analyzed and compared with multi-channel results.

The reconstruction of a moving object from image sequences corrupted by strong structured noise is one of many signal processing problems that require the estimation of the components of a composite field from its sum. In these problems, the composite field is the superposition of one signal that corresponds to the object of interest, other signals that represent clutter, the effect of the transmission medium on the received signal, and signal noise. The traditional solution to these problems is based on modeling the distortion as either a periodic deterministic signal (such as a sine wave) or a random signal (such as Gaussian noise) and using this model to reconstruct the signal of interest. In this paper, we develop a more general algorithm that can be used to reconstruct the relevant object signal from a sequence of frames, with little restriction on the distortion patterns. We begin by developing an analytical model for the reconstruction of composite signals and show that a direct exact solution for this problem does not exist. Next, we apply an iterative procedure based on Kaczmarz's method of image restoration. We show that when the signal of interest has limited support (i.e., is nonzero over a window that is smaller than the image size), an exact restoration is feasible. For this case, we study the effect of the number of frames and the relaxation coefficient used in the iterative algorithm on its rate of convergence. Finally, we discuss other applications of this technique such as the restoration of composite vector fields.

A new form of adaptive median filtering that suppresses impulsive noise while preserving image detail and thin lines is presented. This is accomplished by first generating a binary mask of the image which is subsequently used to drive the type of filtering within the filter windows. The decisions used are either 2-D or directional 1-D median filtering, depending on line structure information from the binary mask. Filtering along the direction of a line preserves the line and suppresses impulsive noise. Experimental results with a real image demonstrate the performance of the algorithm.

The motion of industrial robots generates various risks for equipment and personnel within the workroom. This contribution shows how an automatic visual surveillance system could be used to prevent collisions between industrial robots and operators. We propose a real time system that is able to detect an intruder in a dangerous area, even when there are disturbing illumination changes in the considered shopfloor. The analysis of the sign of the derived image seems very robust. The research has been validated on a site with an industrial robot.

We present the architecture of a system, based on a PC, covering real-time video processing by integrating live video and computer graphics. The video processing and integration is realized by generating and evaluating an on-line MIPmap video texture. This new application of a known filtering technique for still textures in the field of live video processing permits an easy and high performance integration of video and graphics and speeds up the video processing. The evaluation is done by mapping the video texture on surfaces defined by graphics primitives. The graphics primitives are processed by a standard graphics pipeline with special hardware support for the rendering part.

The techniques of digital computer image processing have been employed to investigate the characteristics of paintings by the Italian master, Raphael. In particular spectral histograms have been generated for several of his paintings for individual areas such as hair, face, and garments. It was found that the palette employed by the artist does exhibit regularities. Similarly, spatial frequency histograms of like areas of his different paintings also show certain uniformities. Thus, the first steps have been taken in developing a `fingerprint' for the chromatic tonality and impasta characteristic of this artist's hand. Such information may be of use in trying to attribute controversial paintings where historical data are lacking and chemical data are ambiguous. This trial `fingerprint' has been applied to an unattributed Italian painting known as the `Madonna della Divino Amore.' Although the match to the `Raphael fingerprint' is not perfect it does offer sufficient encouragement to suggest that further investigations into the nature of its materials and its history are warranted.

In the late 19th Century the Old Executive Office Building (OEOB) was constructed adjacent to the White House. Over the years it has housed many of the departments of the executive branch of the U.S. Government. By the turn of the century most of the original painted designs on its walls had been covered over with various `institutional' wall paints. In recent years there has been a move to restore may of the private and public rooms in the OEOB to their original appearance. In the case of the Office of the Vice President during a brief period when the office was vacant, the overpaints were removed in small areas to reveal glimpses of the original wall designs. These design fragments were then photographed and digitally scanned. The techniques of digital computer image enhancement were next employed to clarify, extend, replicate, and rotate these designs. An image of the office was then restored to show its original appearance. Subsequently, the office itself was restored. Before and after images of the office are presented along with the computer-generated reconstruction that guided the restoration work.

In satellite analysis, it is often necessary to determine the attitude of a satellite from a two- dimensional image of the satellite. This paper describes work performed to determine the attitude of a satellite relative to a reference position. The attitude is described in terms of the scaling, rotation and translation of the reference that will result in a pose that can be projected to form the image of interest. The techniques used are based on finding coefficients for the linear combination of basis images and decomposition of a composite transformation matrix. A special representation of the basis images allows the extraction of the linear combination coefficients and the translation parameters in a single matrix multiplication using only three basis images. The methods were initially modeled using a high level mathematics package and applied to simple objects, then implemented on a silicon graphics workstation and applied to a model of an actual satellite. An analysis was performed to determine the sensitivity of the attitude estimation system to input errors and convergence tolerance. The satellite model used and the software tools are available via anonymous FTP.

The advent of symmetrical multi-processor workstations has made it possible to perform computation-intensive image processing on a desktop system with performance approaching that of a large-scale parallel supercomputer. Discussed is the implementation and software architecture of a system developed for use in the processing of infrared spacecraft data. The software constructs infrared scan data into high fidelity images while simultaneously separating celestial point sources from the diffuse background. The methodology relies on parallelism at the operating system level for multiple data element access, and at the source code levels for multiple instruction performance. All modules operate simultaneously on shared memory segments using interprocess communication to improve the data flow through the various required computations.

Texture is an important characteristic of analyzing images. A variety of texture features have been proposed for texture discrimination, whereas a best set of texture features never exists. This paper considers statistical textures which can be viewed as realizations of some stochastic processes, or viewed as images containing no apparent objects. We propose using singular value decomposition (SVD) strategy for texture analysis including (a) using the proportion of dominant singular values of an image matrix as texture features for texture discrimination, (b) the singular value decomposition automatically provides a compression technique for textures due to the dependency of neighboring pixels, and (c) an algorithm based on SVD is proposed to synthesize textures. The texture features derived from SVD are stable according to the stability of SVD. Experiments for discriminating synthesized textures and natural textures, for compressing texture data and for synthesizing textures are also given to demonstrate the proposed strategy.

Devices called fishways or fish passes are constructed in rivers to help migratory fish get over obstacles (dams). There counting windows are used to monitor fish passage by video-based counting. Our goal is to design and construct a vision system to automate this process. Images are taken by a video camera fitted with an electronic shutter in a backlit fishway. They are stored on optical disks in real time but are processed in delayed time. Faced with high volumes of data, a compression is necessary and an electronic board has been designed to accomplish it in real time. The coding method used is based on a run description of binarized images. Then, a tracking process is implemented on a micro-computer to count the fish crossing the pass. It includes fish recognition, which is based on a Bayesian classification process. In order to reduce processing times, recognition operations (labelling, parameter extraction) are accomplished on coded images. Classification results are satisfactory and are improved by the temporal redundancy generated by the tracking process. Image processing time permits the user, on average, to process images faster than they have been stored. Thus there is no data accumulation. At the end of the processing it is possible to edit a result file, to choose a fish, view its crossing images and change its species if wrong.

The advanced remote sensing simulator (ARSIM) is a multi-purpose, interactive satellite sensor simulation environment under development at Martin Marietta AstroSpace, with the cooperation of GE's Corporate Research and Development. ARSIM accepts as inputs (1) an input image at a higher resolution than the desired output, (2) a digital elevation model (DEM) for the terrain in the scene, (3) parameters governing the orientation and attitude drift of the sensor, (4) optical and radiometric properties of the sensor platform and sensor interface parameters, (5) atmospheric conditions, (6) cloud models, (7) targets to be inserted in the image, and (8) types of data processing. Using these inputs, ARSIM generates a high-fidelity simulated image, in the appropriate bandpass, as seen from the satellite. The simulated image incorporates the effect of satellite motion, earth's rotation, jitter, drift, and other motions, optical blurring, atmospheric distortion, occlusion because of clouds and terrain relief, and various other effects. ARSIM allows the user to validate initial specifications on the platform and payload, pose what-if queries, and experiment with the parameters of the satellite. The objective of this paper is to describe the functions and software architecture used in ARSIM.

In this paper, we describe an intelligent stroke extraction method based on modified polygonal approximation. Since a Chinese character is written stroke by stroke, we can regard Chinese characters as a collection of strokes. Therefore, if the strokes are extracted properly, the complexity for recognizing handwritten Chinese characters will be simplified. Using a thinning algorithm to find out the strokes of an input Chinese character has been suggested by many proposed papers. However, the thinning result is very sensitive to noise and the thinning process is time-consuming. Hence, a method for extracting strokes with robustness to noises and time-saving is highly desired. In this paper, we propose a stroke extraction method based on modified polygonal approximation. This method is more efficient and accurate in extracting strokes than any others.

In this paper, a segmentation method of knowledge-radicals for on-line handwritten Chinese characters (OLHCC) is proposed. Using the methods of finding local minimum and finding minimum of sum in local regions, some segmentation lines are obtained. We use a trick called `line segment shortening' to improve the above mentioned methods if overlapped radicals exist in a character. Based on the common writing habits of people, the decision algorithms are proposed to identify the correctness of segmentation lines. Our experimental results are conducted on ten databases of 5401 frequently used Chinese characters that users wrote according to their habits. An average suitable segmentation rate of more than 94% has been obtained, which shows that our algorithm is reliable.

This paper documents a target acquisition and retrieval problem for a mobile robot. The robot's goal is to grasp a known object whose location is unknown. The robot achieves its goal through a visual search followed by physical movement toward the object. To avoid the computationally expensive, general problem of finding a specific object in a cluttered scene, the robot restricts its visual search area to those places that exhibit colors found on the object. When the object is outside grasping range, precise object world-coordinates are unnecessary for the robot to approach the object. Rough coordinate estimates are sufficient if they are quickly computable and improve with decreasing range. As an example of this problem, a robot is programmed to find and approach a specific soda can at an unknown location in a cluttered environment. Color, in this situation, is a more reliable cue to the location of the can than other features. This paper presents a focus of attention algorithm using color, which provides rough estimates of object position. The algorithm described here is related to the histogram backprojection algorithm of Ballard and Swain, but it does not require the object image size a priori.

We develop a new morphological hit-miss algorithm for detecting the features of binary images. The standard hit-miss algorithms use block structure patterns for matching features but in many practical situations this may cause problems because the size of the object is usually unknown. In the proposed algorithm we match the corners of the object instead of the block structure pattern -- this approach works nicely because most objects have different structure elements for the corners. The size of the object plays a less significant role in our recognition algorithm than in the standard algorithms. Our method can be easily adapted to recognize the features of different types of objects. We have implemented the proposed algorithm to recognize the muzzles of guns in military vehicles.

In security monitoring, the false alarm rate may be reduced if intelligent reasoning is applied to moving regions identified in each camera. Stable objects of appropriate size, position, and velocity may be allowed to trigger alarm. In the following, moving regions are identified against a stationary background by thresholding accumulated image differences. Tracking allows the frame-to-frame feature correspondences to be established, and the feature motions and covariances to be updated. Regions extracted in this manner are often not stable for more than a few frames. In order to cope with this, a two level hierarchical scheme is introduced in which groups of features of consistent 2D motion are grouped into moving regions of interest (MROI) each attributed with their own 2D motions. A hierarchical matching scheme is then employed to guide the matching of MROIs and regions. Such an approach provides the spatial and temporal stability absent at the moving region level. In addition such high level features are the appropriate level at which to reason about image events in the context of security.

This paper presents a method for the classification of textures using quadrature mirror filter (QMF) bank subband decomposition in combination with statistical descriptors. In our combined method the QMF bank splits the input image into four subbands, and statistical descriptors based on co-occurrence matrices are computed from the subsampled low-low band. The experiments demonstrate that the combined method has better classification performance than that of statistical descriptors computed from the co-occurrence matrices of the whole texture image. In addition, the experiments demonstrate that the combined method based on computationally efficient IIR QMF banks yields approximately the same classification results as the combined method based on classical FIR QMF banks.

In this paper we present a technique to design 2-channel filter banks in 3 dimensions where the sampling is on the FCO (face centered orthorhombic) lattice. The ideal 3-D subband is of the truncated octahedron shape. It is based on the transformation of variable method and it is equivalent to the generalized McClellan transformation. The filters are FIR, have linear phase, and achieve perfect reconstruction. Although the subband shape is quite complicated the ideal frequency characteristics are well approximated. This is illustrated with an example. The technique provides the flexibility of controlling the frequency characteristics of the filters with ease. The filters can be implemented quite efficiently due to the highly symmetrical nature of the coefficients of the transformation.

A 3-D tool wear analysis method has been developed to measure erosive wear parameters such as area and volume automatically and objectively. Parallel light stripes are projected onto a wear area at a known angle. Knowledge about the light pattern is utilized in processing the digitized image data to compute local depth. The new method was successfully applied to sub- optimal images where magnification is necessary, such as cutting tool wear. Since this is a general image processing problem with a structured light source, this method can be easily adapted to other applications.

In this paper, a programmable accelerating chip with parallel and pipelined computation capabilities for optical Chinese character recognition (OCCR) is presented. The chip contains eight matching processing elements working in parallel and a memory-saving sorter which can choose the best 32 candidates. The structure of the design is flexible and modular, which is capable of processing different start points and end points of the template feature. Fabricated in 1.0 micron CMOS gate array, the chip contains approximately 16,000 gates. It has been tested to be fully functional. By using this chip, a powerful optical Chinese character recognition system with high speed, high recognition rate, and accumulated learning ability is developed. From the experimental results, the process speed of this chip can be up to 200 characters per second, which is one hundred times the conventional pure software process speed. The chips can be used in parallel for a larger template character system and will not increase the process time.

We describe a new method for computing approximations to the marginal probability mass function of the random variables in a Markov random field (MRF). When applied to the a posteriori MRF, this yields approximations to the conditional marginal probability mass function, which is the key quantity in a Bayesian classifier. We apply these ideas to an optical agricultural remote sensing problem where they outperform the pixel-by-pixel ML classifier by 38%.

This paper presents the main results obtained through a study on aircraft identification and attitude estimation conducted by Thomson TRT Defense for the French Ministry of Defense/Direction Generale de l'Armement/Direction des Constructions Aeronautiques. The purpose of this study was automatic assistance to aircraft identification. Indeed, modern fight airplanes are equipped with optronic systems capable of detecting and tracking enemy aircraft. In order to react quickly, the pilot must know at least the target type and possibly its identity. Recognition of the target type and attitude is obtained by matching the observed image with patterns belonging to a database. Two matching algorithms, which have been tested, are presented. The first one, based on the contour Fourier transform, needs the complete target silhouette extraction. The second one, belonging to the class of prediction and verification algorithms, compares the individual parts of the target to the database and is able to recognize the target, even when it is partially occluded or ill-segmented due to the lack of contrast between the target and its environment. An original feature of the algorithm stays in a validation process which increases the reliability of transmitted answers. In case of low confidence, no answer is provided. In addition, successive answers are consolidated. This strategy is interesting especially for image sequences where the tracked airplane achieves attitude evolution or even simply flies over various backgrounds. The main output of this study is the parametric analysis of various factors which influence performance such as contrast, background complexity, distance, attitude and type. The evaluation method, largely based on image synthesis (including image sequences), allows fine interpretation of statistical results. Misclassification errors occur when resolution is not sufficient or when complex backgrounds cause erroneous segmentation. Best results are obtained when the object image is large enough and presents good contrast with its background. The results confirm the major interest for range information in order to discriminate big transport aircraft from fighters. Finally, image sequences show answer quality increases along with the time, in attitude as well as in type.

This paper deals with the fusion of data delivered by an active LIDAR (light detection and ranging) sensor and a passive FLIR sensor, in the framework of ground-to-ground automatic target recognition (ATR). ATR operators usually analyze features which are measurements of the observed object. Recognition process efficiency highly depends on both variety and stability of these features. Active/passive fusion clearly provides a large amount of miscellaneous data. It is shown here that it can also greatly improve target silhouette extraction and thus increase the stability of morphological features. Target silhouette extraction is performed by an image segmentation algorithm which analyzes LIDAR and FLIR registered images. The main stages of the algorithm are described. Experimental results obtained by processing outdoor scene data are presented.

One of the main problems faced in the development of pattern recognition algorithms is assessment of their performance. This paper describes the development of a novel technique for the assessment of information content of 2-D patterns encountered in practical pattern recognition problems. The technique is demonstrated by its application to multi-font typed character recognition. In this work we first developed an information model applicable to any pattern, and its elaboration to measure recognition performance, and second we used this model to derive parameters such as the resolution required to distinguish between the patterns. This has resulted in a powerful method for assessing the performance of any pattern recognition system.

We develop a model-based vision approach to estimate the `aspect angle' of a target in a forward-looking infrared (FLIR) image. A set of 3-D models of military vehicles is created using a CAD/CAM package. From this set, a database of 2-D images is created by rotating the models and then projecting onto a given plane. A matching algorithm is then applied to match a signature of the given target with the images in the database. We use an algebraic approach to represent images which has the advantage that a big speedup is possible if the fast Fourier transform (FFT) is applied to compute the polynomial multiplications involved in the algebraic approach.

Heursistics for programming real-time color recognition hardware, controlled by a symbolic (Prolog) processor, are described. The paper addresses the engineering aspects of color recognition and its application to such tasks as industrial inspection and sorting. The techniques discussed are heuristic, not algorithmic; they are not particularly well suited to those applications where there are subtle color changes. However, in numerous experiments, they have been found to produce useful results, in a wide range of important industrial inspection and sorting tasks, where relatively coarse color discrimination is sufficient. The application of these techniques to the inspection of standard multi-color packaging materials is highlighted.

Techniques for the reconstruction of quadric surfaces (cones and cylinders) from disparity measurements obtained by a stereo vision system are presented. For the initial estimation of the geometrical parameters of cones, a regression based approach is adopted from [Newman, Flynn, and Jain, CVGIP: Image Understanding, Vol. 58, pp. 235-249]. A new method based on surface normal estimation is developed for the cylindrical case. For the refinement of the parameter estimates, the Levenberg-Marquardt nonlinear least squares algorithm is used as in [Flynn and Jain, Proc. 1988 IEEE Comput. Soc. Conf. Computer Vision and Pattern Recognition, pp. 261-267] but two mistaken formulas (one for a cone and the same for a cylinder) are corrected. The prior approaches for quadrics in stereo vision have been more or less approximative. Our contribution is to fit the exact quadric shape to disparity data using the existing methods for range data. The disparity space offers a convenient frame to analyze the differences between the measurements and the reconstructed model.

Some performance measures for partitioning images among hypercube connected processors are presented. The paramount effect of row-major ordering of image bytes is explicitly taken into account. Subimages are split at row boundaries first and downloaded over a spanning binomial tree. Subimage nearest neighbors are mapped to processor neighbors. A theorem which indicates that subimage locality is preserved is given. Practical constraints of a real machine (nCUBE 2) are incorporated. Performance comparisons between this and related image communication techniques are presented.

In this paper the problem of detecting objects in the presence of clutter is studied. The images considered are obtained from both visual and infrared sensors. A feature-based segmentation approach to the object detection problem is pursued, where the features used are computed over multiple spatial orientations, and frequencies. The method proceeds as follows: A given image is passed through a bank of even-symmetric Gabor filters. A selection of these filtered images is made and each (selected) filtered image is subjected to a nonlinear (sigmoidal like) transformation. Then, a measure of texture `energy' is computed in a window around each transformed image pixel. The texture `energy' features, and their spatial locations, are inputted to a least squared error based clustering algorithm. This clustering algorithm yields a segmentation of the original image -- it assigns to each pixel in the image a cluster label that identifies the amount of mean local energy the pixel possesses across the different spatial orientations, and frequencies. This method is applied on a number of visual and infrared images, every one of which contains one or more objects. The region corresponding to the object is usually segmented correctly, and a unique set of texture `energy' features is typically associated with the segment containing the object(s) of interest.

In this paper, a novel idea of preprocessing which is based on features to perform a global noise reduction is proposed. First, we extract the stroke features of Chinese characters and define contour data structure to describe it, and then, by means of clustering and classification techniques, we can get templates to proceed modification. This method is very different from traditional treatment in preprocessing. The experimental tests of 5401 handwritten characters show the satisfactory results and support that it is an available method.

Methods of 3-D visualization of the brain based on fuzzy c-means (FCM) classified magnetic resonance (MR) images and a neural network trained on the FCM data are presented. A 3-D MR scan of a volunteer serves as the basis for the unsupervised classification techniques. The images were first classified into different tissue types by using FCM. The classified images were then reconstructed for 3-D display. Results show that individual tissue types can be discriminated during the 3-D rendering process. A neural network trained on the fuzzy classification data was also implemented. By using the cascade correlation algorithm during the network training, much of the tedious training work was avoided. The preliminary results from the neural network approach are quite encouraging.

A new low cost imaging system has been devised to detect and measure joint movement to help with the diagnosis of ligament injuries in the human knee. The system uses a domestic video camcorder to record the movement of marks on a patient's knee as it is flexed. The pictures are then fed into the imaging system, where the coordinates of each mark are determined for each angle of flexion. The coordinate data is then processed to show the dynamic operation of the knee, from which an assessment of ligament damage can be made. The imaging system is comprised of a PC host, a commercial frame store, and a custom built TMS320C40 digital signal processor (dsp) board. The dsp is used to perform correlation and other imaging functions, to automatically determine the mark coordinates in real time. This paper describes the application and development of the system, and gives the results of the research to date.

Current computer graphics techniques can generate 3-D views of the human anatomy from magnetic resonance images. These techniques require that the images first be segmented into the various tissue types. However, there has been no fully automated system that can perform this task on a single set of high-resolution 3-D magnetic resonance images. We present a fully automated segmentation algorithm based on the 3-D difference of Gaussians (DOG) filter. A novel method for the classification of regions found by the DOG filter, as well as a correction procedure that detects errors from the DOG filter, is presented. Regions are classified based on the mean gray level of the voxels within closed contours. In previous work, the user had to manually split falsely merged regions. Our automated correction algorithm detects such errors and splits the merged regions. Spatial information is also incorporated to help discriminate between tissues. Encouraging results were obtained with an average of less than five percent error in each image. Integral shading is used to obtain a 3-D rendering of the data set.

A generalized adaptive median filter (GAMF) is introduced for more robust noise suppression and edge preservation in computed tomography. The GAMF employs a set of adaptive linear filters in conjunction with an adaptive median filter. Each linear filter is designed to preserve a feature in one orientation. The window sizes of these filters are adaptive to the local statistics. The output of the GAMF is selected from the linear filter output set that is closest to the adaptive median filter. This filter has been successfully utilized in the computed tomography to combat severe streaking artifacts resulting from excessive x-ray quantum noise. Its effectiveness has been demonstrated in phantom tests and clinical studies. In addition, little impact on system resolution can be observed.

For the evaluation of cardiac morphology and cardiac dynamics from a time series of angiograms, reliable and precise extraction of coronary arteries in each image is the key technology. An algorithm to detect vessel contour with high reliability is developed using the characteristics of coronary artery motion. The movement of coronary arteries is relatively larger than that of background tissues, so we can get high contrast vessel contour by substraction between two time different spatially differentiated images. This spatiotemporal operation gives us better coronary artery detection in the presence of noise than other spatial operations by using accumulation of spatiotemporal subtraction images for the interval where the background movement is negligible. Detected coronary artery branches are described with blood vessel segment sequence for the right and left angiogram. These two vessel structures are compared and the correspondences found. Thus 3D coronary artery structure is determined and displayed on the CRT. The effectiveness is discussed on actual dynamical coronary angiogram compared with spatial operation on the time frozen single image. Our algorithm works satisfactorily in this example.

An image analysis method of measuring the effectiveness of a toothbrush in reaching the interproximal spaces of teeth is described. Artificial teeth are coated with a stain that approximates real plaque and then brushed with a toothbrush on a brushing machine. The teeth are then removed and turned sideways so that the interproximal surfaces can be imaged. The areas of stain that have been removed within masked regions that define the interproximal regions are measured and reported. These areas correspond to the interproximal areas of the tooth reached by the toothbrush bristles. The image analysis method produces more precise results (10-fold decrease in standard deviation) in a fraction (22%) of the time as compared to our prior visual grading method.

The human recognition probabilities of blurred rotationally symmetric shapes have been studied using computer generated images displayed in a 128 X 128 pixel area on a TV monitor. The shapes employed included a series of regular polygons, crosses, and rectangles. These were blurred by convolution with two dimensional Gaussian functions which had standard deviations ranging from 0 to 29. Images of the blurred shapes were presented to observers in a random order and with a random extent of blurring. After each presentation the observer decided which of the shapes was most likely to be represented by the image displayed on the screen. A correlation has been found between the extent by which a shape may be blurred before it ceases to be recognizable and the difference between the original shape and a circle of the same area. This correlation has been expressed as an empirical relationship between the probability of recognition and the standard deviation of the Gaussian blurring function when the latter is normalized by a function which depends on the original shape and the one it is being confused with. This relationship has been applied to a series of irregular shapes to predict the amount of blurring required before they too cease to be recognizable. These predictions have been compared to experimental observations for the irregular shapes considered. The probability of recognizing an image may be used as a measure of image quality. The empirical relationship derived from this work could, therefore, form the basis of a new objective performance measure for thermal imaging systems.

In this paper, new techniques of determining the focus of a disease position in a gamma knife operation are presented. In these techniques, the transparent 3D color image of the human body organ is reconstructed using a new three-dimensional reconstruction method, and then the position, the area, and the volume of focus of a disease such as cancer or a tumor are calculated. They are used in the gamma knife operation. The CT pictures are input into a digital image processing system. The useful information is extracted and the original data are obtained. Then the transparent 3D color image is reconstructed using these original data. By using this transparent 3D color image, the positions of the human body organ and the focus of a disease are determined in a coordinate system. While the 3D image is reconstructed, the area and the volume of human body organ and focus of a disease can be calculated at the same time. It is expressed through actual application that the positions of human body organ and focus of a disease can be determined exactly by using the transparent 3D color image. It is very useful in gamma knife operation or other surgical operation. The techniques presented in this paper have great application value.

In this paper, a new method of extracting the configuration of a three dimensional object is presented. It can be used in computer vision or robot vision. In this method, two or more pictures of a three dimensional object are taken and stereo image pairs are composed of these pictures. The intrinsic and extrinsic parameters of camera are solved using these pictures and then the images of every stereo image pair are matched using a new method presented in this paper. When the coordinates of point on left or right picture in image space are given, the coordinates of corresponding point on right or left picture can be calculated automatically. The coordinates of point on a three dimensional object in object space can be solved using the coordinates of stereo image pair. After this process has finished, all the coordinates of the three dimensional object in object space are obtained. The configuration of the three dimensional object has been extracted. We use this method to extract aircraft configuration. The result is satisfactory. It is expressed through actual application that the method presented in this paper is correct, the result is reliable. It is very useful for computer or robot vision. The method presented in this paper has theoretical and application value.

The most challenging question in image restoration nowadays is how to obtain blur information from the degraded image. Specifically, the identification of PSF has been a subject of great interest. However, many early identification methods are, in general, more expensive computationally, and have some limitations, more or less. In this paper, a new identification and restoration method based on the power spectrum filter was proposed by means of equivalent power spectrum assumption. Specifically, a recognition method, for the power spectrum of the original image in the filter transfer function expression, was suggested through exploring spectra of the degraded image. The major merits of this method are it is simpler, more practicable, and more economical computationally than the others, such as identifying PSF parameters.

Sometimes the classifiers based on the features extracted from patterns may not be robust, in this case, to obtain better classification results, man's interruption is needed, then subjectivity and uncertainty due to man's action are followed as a result. In this paper, an algorithm able to automatically create a classifier is provided by the technique of learning from examples, with which pattern recognition, such as the facial images recognition, are completed.

A new method, based on the binary circular harmonic function of target pattern, is developed for rotational invariant pattern recognition. The computer simulation results show that this method yields a much higher correlation peak and a better correlation discrimination ability compared with the method of classical circular harmonic component.

The Fisher linear discriminant vector has been used as the optimal linear method in solving pattern classification problems. This paper proposes an iterative algorithm to calculate a global optimal set of discriminant vectors under the global Fisher discriminant criterion. The main advantage of our algorithm is that the scatter matrices in the subspace spanned by all discriminant vectors in the proposed optimal set have the global minimum within-class scatter and global maximum between-class scatter as compared to the Foley-Sammon local optimal set.

The model-based image coding might be the potential method for very/ultra low bit rate visual communications. However, some problems still remain for video practice, such as a finer wireframe 3-D model, precise rule for facial expressions analyzing, and automatic feature points extraction for real time application, etc. This paper proposes a feasible scheme of model-based image coding based on a deformable model which would be suitable for very/ultra low bit rates transmission. Meanwhile, some key techniques, such as automatic face feature point extraction based on a priori knowledge for real time applications and the method of AUs separation of a face on various expressions, is given.

In this paper, we introduce a neural network recognition method, MENN (minimum error neural network) method, in target recognition. From the target gray sequences, we can extract some useful characteristics. Then we use these features as the input data of the MENN classifier. By these characteristics, using the MENN classifier we can easily pick out the true targets from the candidate target sequences. MENN recognition method can not only pick out the true target and reject the false targets, but it also gets rid of the baits. Therefore, it has high reliability. Moreover, it has many advantages, for example, its training is a one pass process, its test process is not only simple but also straightforward, and its calculation is simple, etc. On account of those advantages, MENN recognition method is adaptive to the need of realtime processing.

This presentation discusses the problem of segmentation of nuclei in cytological color images in different color spaces, namely RGB and HSI color spaces, for the detection of lung cancer cells. For the segmentation in each color space, the background and foreground of the images are first defined, and the chromatic mean values of the background and foreground are then extracted. In the learning phase, based on the chromatic mean values of the background and foreground of training samples, an adaptive threshold function is constructed for each color space using the B-Spline technique. The nuclei are then segmented by thresholding using the adaptive threshold function obtained in the learning phase. Comparisons between the segmentation in RGB color space and in HSI color space are carried out.

A special technique for digital polarization images treaty is presented in detail. This technique is used for Earth surface remote sensing purposes and deals with the full Stokes vector analysis of light reflected from surface. The underlaying surface is modeled as a little anisotropy absorption medium. The accuracy characteristics of the proposed method on the base or real digital images are presented. Specific limitations of technique for polarization digital images treaty in visible are considered. The ability of this technique usage to SAR polarization images treaty is also discussed. The main requirements for obtaining radar polarization images are proposed.

Subband coding is becoming a viable compression method for video transmission, since the subband system can be efficiently split into a base layer and an enhancement layer without incurring any overhead. This is desirable in ATM and other packet networks. The base layer can be a standard coding algorithm such as H.261 or MPEG. In a subband-based video coding system, the higher-frequency subbands contain a small amount of the signal energy, but a large fraction of the subband data. Coarsely quantizing this higher-frequency information can achieve the desired bit rate, but will lead to lower resolution since the high frequency edges are perceptually important. In this paper, methods are studied to minimize the distortion of the coarsely coded higher frequency subbands through the use of entropy-constrained trellis coded quantization (ECTCQ) and entropy-constrained vector quantization (ECVQ).

Retinal exudates are a common manifestation of damage of the human eye ground. An adaptive threshold algorithm was developed to detect and to measure the exudates in gray value images of patients with diabetic retinopathy. The images of the eye ground were acquired by monochromatic illumination of a scanning laser ophthalmoscope. By means of a dynamic thresholding procedure, we are able to omit the common usual preprocessing shading correction since the compensation of irregularities of illumination is implicitly contained in the algorithm. In order to get an objective assessment e.g. of the course of a disease or the success of a therapy we take into account the area as well as the location of the retinal exudates. The analysis is operator independent except for the choice of an area of interest.