A bandwidth-compression scheme is presented based on spectral and spatial data correlations that can be used to transmit radiometric data collected by sensitive high-resolution sensors. The Karhunen-Loeve transformation is employed to remove spectral correlation, after which the data are treated with an adaptive discrete-cosine-transform coding technique. Coding errors are spread over entire individual data blocks after reconstruction because the coding is done in the transform domain. The technique reduces bandwidth while maintaining near-lossless coding, the ability to handle a high dynamic range, and the establishment of a maximum-coding-error upper bound. The approach is capable of some feature-classification capability, and each image is a blend of data from the entire set of spectral images. The method for bandwidth compression therefore permits the evaluation of a range of information without examining all images in the data set

An algorithm for segmentation of cell images and the extraction of texture textons based on vector quantization is presented. Initially a few low dimensional code vectors are employed in a standard vector quantization algorithm to generate a coarse code book a procedure which is equivalent to histogram sharpening. Representative gray level value from each coarse code vector are used to construct a larger fine code book. Coding the original image with the fine code book produces a less distorted image and facilitates cell and nuclear extraction. Texture textons are extracted by application of the same algorithm to the cell area using a larger number of initial code vectors and fine code book. Applications of the algorithm to cytological specimen are presented.

A new efficient computational technique for digital itiage bandwidth xxtLpression is presented. This technique is based upon an improved variation of the Gercbberg-Saxton--Papoulis (GSP) Fourier phase reconstruction algorithm. A digital image is converted into a ccitiplex 2D signal in the spatial dcxtiain by imposing a known randan phase angle. This step is then treated as a projection onth a closed convex set of 2D signals with the sait kncn phase angle. Using the theorns of projections, a functional j.iation is fornd which includes, together with the standard constraints of the knn phase angle in the Fourier dmain and the cotract image spatial support, a new orthogonal projection operator with its relaxation parameter. Thus formed, the new iterative algorithm produces exact image reconstruction in a snall number of iteration steps. The new algorithm is applied to a partially observed Fourier phase image to prcxluce the original image reconstruction of good quality. Fuftherixre, multiple oversampling of the Fourier phase in the frequency donain is substantially reduced. Several experimental results are presented.

Two types of frequency characterization of the discrete cosine transform (DCT) have been analyzed in detail. One is performed in the natural spatial frequency domain and the other in the eigenspace of the firstorder Markov stationary random process. In the past the direct conversion from Fourier transform toDCT has been very difficult. It requires either doubling of the input data or reshuf fling the input data sequency. In this paper we have derived a unitary transform that allows one to directly convert a Fourier transform of natural sequence input into a DCT. Furthermore though it is known that the DCT asymptotically approaches the KarhunenLoeve transform (KLT) of the firstorder Markov stationary random process no exact relationship between these two transforms has been given. This paper derives the exact relation and exhibits the frequency characteristics of a DCT. Applications to image data compression and enhancement are also included.

Discrete cosine transform (DCT) coding has been emerging as a key element for image data compression. Conventional DCT coding algorithms which treat all the image areas indiscriminately unfortunately give nonuniform image quality for various image contents. This motivates work on DCT schemes adaptive to the image contents so that a better tradeoff between bit rate and image quality can be achieved. In this paper a DCT algorithm with adaptive thresholding and quantization is combined with variable block size segmentation to further improve the coding performance. A new segmentation criterion is proposed. Some simulation results are given to illustrate the superiority of this adaptive DCT algorithm with segmented blocks. It is also shown that this algorithm poses itself as a promising compression method to deal with the composite images consisting of text/graphics and natural scenes.

A data structure based on the Gray code is presented which while maintaining the computational efficiency of linear quadtrees requires less storage memory. Unlike linear quadtrees where one can condense only to squares of 2hierarchically code can condense the adjoining black nodes to rectangles of the size 2'' x 2 . Due to this data can be condensed to a greater extent using Gray code. Algorithms for condensed encoded Gray code structure are similar to those of linear quadtrees. This paper presents algorithms for Gray code encoding comparing members adjacency and neighbour finding condensing data component labelling area and perimeter calculations.

Since 1975, starting with the fundamental work of J.O. Drewery ( 1 ) , the problem of clean PAL decoding was deal t wi th the tool of multidimensional filtering. The need to fulfil subjective requirements led the authors of the previous works to design very simple filters with poor frequential characteristics, obtained without software design tools. The aim of this work is to show that software tools for the design of multidimensional (MD) filters using linear programming are to be considered powerful also in solving problems of luminance/chrominance (Y/C) separation in received PAL signals. Furthermore this approach allows to harmonize the design specifications and the requirements imposed by the human eye.

An image representation infrastructure termed the macro-microimage representation is proposed in this paper. The representation uses the macroimage for the feature such as velocity and the microimage for the preservation of spatial relations. This architecture is in favor of local connections and is suitable for some motion analysis tasks. The mapping of this infrastructure to a multi-processor parallel architecture is also discussed

By generating Fourier descriptors based upon the waveform induced by a pattern''s projection a number ofclassic difficulties with the Fourier-descriptor methodology are mitigated. Not only are the descriptors invariant with respect to scale translation and rotation (as is usually the case) they are also continuous in the Hausdorff metric and robust with respect to both point noise and occlusion insensitivity with respect to minimum occlusions being perhaps their most significant advantage. Continuity in the Hausdorif metric allows prediction of the effect on the descriptors when morphologically filtering a pattern. The effect of occlusion is also predictable. An additional advantage is that they can be computed relative to a thresholded image without first finding an edge thereby avoiding the difficulties typically present in thinning and orientation determination.

A model-based image-analysis technique is presented which allows the calculation of combustion and flame parameters and physical quantities. Emission-tomographic techniques are employed to determine the radiation-intensity distribution of the flame, and only one image is required in the case of rotational symmetry. After regularizing and constraining the radiation-intensity distribution the intensity distributions of two or more images are examined at different optical wavelengths. The image analysis can be used to determine 3D temperature distributions and gas concentrations for each pixel with pyrometric methods, and the results compare favorably with other measured temperatures. The technique can be used to analyze particular radiation characteristics of combustion processes by employing optical filters

This paper briefly describes the stochastic image model and the image analysis technique for X-ray CT, and then validates the use of this model and technique to MRI modality. The results of image analysis from the use of X-ray CT and MRI are included to show the promise and the effectiveness of the developed technique.

The integration of a vision system in robotics imposes the constraints of real-time. The choice of the representation model in addition to the object machine is then determining. The data to be processed being considerable our practical approach consists of making the access to the bulk data easier by its distribution on various active memory units each one having its own checking means. The formalism used is the actors one enhanced by the description possibilities of frames model this allowing a hierarchical symbolic representation of the visual scene thanks to the abstract and concrete links.

Performance of any gradient type edge detector in the presence of noise critically depends on threshold selection at its output. In this paper it is shown that under the assumption of white additive noise with normal distribution at the input noise gradient magnitudes at the output of edge detector have Rayle igh distribut ion. The proposed threshold selection procedure is based on the estimation of a single Rayleigh distribution parameter from gradient magnitude histogram which is proportional to input noise standard deviation. Threshold is then calculated on the basis of the est imated no ise standard deviat ion and predetermined probabi 1 i ty of false edge detection. Using derived relationship between noise gradient distribution and averaging parameter o in direct ional den vat ive edge detector a who le edge detect i on process can can be enhanced by adjust i ng both thresho 1 d and the amount of averaging. Therefore if an additional constraint of minimal edge intensity that has to be detected is given edge detector can adjust the amount of input averaging to match the given false edge detection probability and estimated noise variance with minimal value of averaging parameter o. Proposed technique is demonstrated on several examples.

Most implementations of Fast Fourier Transform (FFT) algorithms available in software packages and libraries require the number of points on the input to be an integral power of two. However most digital images especially those obtained in PC based systems will seldom meet this requirement. This paper will present a simple computational technique to adjust image dimensions to an appropriate size. Nonlinear polynomials are used as the basis for the scheme. The derivation of the basic interpolation functions is presented and a basic three by three mask is obtained. Observations regarding properties of image and mask are made which lead to the reduction of the mask to a three by one. An optimal procedure for utilizing the mask is also presented.

In this paper a specific data classification approach is presented which can be used in coders for transmitting video sequences. The image data is locally classified into four categories prior to the coding process which is then divided in four modes in order to exploit the masking effect of motion and to reach high performance. The four categories are: moving area containing an edge (data class Cf) moving area containing no single edge (class C/j) still area containing an edge (class C) and still area containing no single edge (class CE). This classification approach is based on the fact that the human visual system is especially sensitive to edge information and this in function of the motion behavior. Generally it is known that the edge information is critical to the human eyes. In the moving areas however the human eyes are less critical of the spatial details in comparison to that in the still areas because of the masking effect of motion. The data in the moving areas can be treated relatively coarsely but when edges are observed the quality criterion should still be raised appropriately. Using this four-mode coding method the subjective reproduction quality can be improved even in the case of low bit-rate applications. This classification approach is applied in the conventional motion compensating hybrid (DCT/DPCM) coder although it can also be with other coders.

Motioncompensated interframe prediction has been applied widely in digital television signal coding specially in low bit-rate coding such as videotelephone and videoconference. A key element in motion-compensated interframe prediction is the motion estimation algorithm. In this paper a new motion estimation algorithm is presented. It is based on the block recursive (gradient) method and makes use of some of the advantages of the blockmatching method. Motion estimation with non-integer pd accuracy can be obtained with only a few iterations. In addition some techniques with respect to motion correlation and motion tendency estimation which we proposed before [1] are applied to the estimation algorithm which are efficient enough to improve the performance of the motion estimation and the homogeneity of the estimated motion vector field. The experiment shows that the proposed algorithm has a much higher estimation accuracy and a much better prediction performance than the conventional block motion estimation algorithms. Moreover a coding approach designed effectively to minimize the bit-rate necessary to present the motion vector field is proposed as well. A high compression rate for the transmission of the motion information has been achieved. Finally the simulation results of motioncompensated interframe prediction for low bit-rate coding based on the proposed algorithm and of the motion information encoding are presented.

Vector quantization is an effective spatial domain image coding technique at under 1 . 0 bits per pixel. To achieve the quality at lower rates it is necessary to exploit spatial redundancy over a larger region of pixels than is possible with memoryless VQ. A fmite state vector quant. izer can achieve the same performance as memoryless VQ at lower rates. This paper describes an athptive finite state vector quantization for radiographic image sequence coding. Simulation experiment has been carried out with 4*4 blocks of pixels from a sequence of cardiac angiogram consisting of 40 frames of size 256*256pixels each. At 0. 45 bpp the resulting adaptive FSVQ encoder achieves performance comparable to earlier memoryless VQs at 0. 8 bpp.

This paper presents an improved algorithm for motion compensated image coding. The algorithm can be appied to videophone or video-conferencing. Background detection is used to reduce the bit rate by eliminating the motion information of the block. For a better estimation the object motion some of the blocks which include edge information are divided further into several subblocks according to the edge components. The algorithm shows the improvement of signal -to-noise ratio in the range of 1. 57-5. 57 dB compared to the other existing algorithms. The algorithm also shows the stable behavior for the abrupt movement as a result of the explicit usage of the edge information in the motion estimation.

We are considering a texture as an image consisting of number of different subimages. An attempt is made to findout all(different) subimages present in a given textured image and the number of times such subimages are repeating. By taking the advantage of the Burt pyramid processing algorithm a system has been developed to solve the above problem in real time. The hardware setup is also presented in this paper. In this analysis two dimensional entropy is used o solve the above said problem.

This paper presents and compares different Vector Quantization techniques for image coding. First the Linde- Buzo-Gray (LBG) technique is considered whose effectivness is known in achieving a minimum distortion partition of the training vector hyperspace: the convergency of the seed points toward the centroids of a sb1e configuration is discussed. Two other approaches are then examined: the first one based on Cohonen neural rt"tworks the second one based on a tree structure to reduce the computational burden during the codebook construction and the actual coding. Perfoimance evaluation is made in terms of bitrate and reconstruction distortion. Experimental results and performance curves are also reported concerning application to static image coding.

The extraction of depth information from stereo image disparities using a 2-D power cepstrum technique as well as the usual correlation technique is presented. The edges of the input stereo images were enhanced by a Laplacian of Gaussian (LoG) operator modified for fast implementation1 . The depth of an object surface is calculated from a simple nonconvergent triangulation technique. The methodology presented in this paper allows a fast yet accurate implementation of computational stereo vision in a machine vision system. The basic concepts of the present paper have been further extended and a motion stereo model has been developed for reconstruction of three-dimensional (3-D) geometry of natural objects using time-sequenced images with capabilities of yielding dense depth maps

A system is proposed for a fully automatized analysis and identification of a human face appearing in a picture. The system learns a face by analyzing it and storing the extracted features in a database. This analysis starts from a very robust detection of a face and its position scale and orientation based on the position of the eyes. The. four transformation parameters define the face-specific coordinates which are used during the entire analysis. In these coordinates some face features are detected and stored which belong roughly to three categories: geometrical proportions surface properties and iconic features. The identification of a person using a different picture of a known person also starts from a face analysis. Using the features the presented face is correlated with all faces present in the database. The correlation of the extracted features contains two steps: individual measures (correlating iconic features computing differences in geometrical proportions or surface properties) give the similarity between two faces for one feature while a global measure (a Euclidean norm for instance) combines these similarity values. The system is therefore able to identify a known person by finding the best match. The results of the system on real images are presented. It is shown that the identification is quite selective.

A new iterative technique for frame-to-frame image motion estimation is introduced and impleniented. The algorithm presented in this paper is based on the maximum likelihood criterion and is referred to as the GML algorithm. This scheme requires the covariance function matrix of the motion a priori. For this reason a possible motion model will be introduced and implemented. Simulation experiments are presented which investigate the performance of the algorithm in conjunction with real and synthetic images. Key Words : Motion Compensation Maximum Likelihood Covariance Function Markovian Field.

Filtered 2-D fractionally differenced discrete processes are proposed as texture models. An iterative algorithm is presented for estimating the parameters of the model from a given texture image. Synthesized textures are provided to demonstrate the power of the proposed model.

Texture image which is applied in many fields is composed of two kinds of regions and between them there is a high contrast. Although the texture image may be not have a certain run in the whole image it has a certain run in a small window or local area. By the use of the run information we can easily determine local threshold to segment the image. In this paper a fast segmentation algorithm based on cross threshold selection is presented. Through the use of pipe-line technology both in x- and in y-directions of image plane simutaniously the algorithm is able to not only get a good segmented image but also save much time even in SISD computer.

We suggest an active method for overcoming a discontinuous problem of shape from shading. In general the characreteristics of discontinuity doesn''t well come out in scene. Hence we use a camera with the small depth of field so that the attribute of discontinuity is exposed appropriately. In this case 3D information is represented by the degree of defocusing in 2D image plane. And objects in scene can be segmented according to the degree of defocusing. We simulate four algorithms - Brooks Horn''s algorithm the revised its algorithm the resistive net without line processor and the up and low threshold resistive net with respect to the synthetic image. We can see that the success of algorithm depends on the precision of detecting the degree of defocus.

An algorithm for object extraction in a random field background using a multiresolution pyramid based on a modified weighted pixel linking model is described. The original model'' depends on the geometric distances node-father and on a density distribution function of the node''s sons characteristics for extracting compact objects. A k parameter was additioned to the Gaussian distribution function that allows a better fit for the objects to be recognized. The gray levels standard deviation (s. d. ) of the objects in the original image can have an important role in the segmentation process for selecting the parameters needed to clean the histogram of the segmented image of the natural texture and noisy regions. The results were applied to the extraction of a characteristic defect of a cork agglomerate tile.

Lockheed has designed and implemented a prototype real-time Video Enhancement Workbench (VEW) using commercial offtheshelf hardware and custom software. The hardware components include a Sun workstation Aspex PIPE image processor time base corrector VCR video camera and realtime disk subsystem. A cornprehensive set of image processing functions can be invoked by the analyst at any time during processing enabling interactive enhancement and exploitation of video sequences. Processed images can be transmitted and stored within the system in digital or video form. VEW also provides image output to a laser printer and to Interleaf technical publishing software.

A method for reliable recognition of fluorescent stained cell monolayers has been developed. This method utilizes convolution filtering and object intensity dependent thresholding to achieve accurate recognition. These operations are performed rapidly by dedicated image processing hardware. Details of the transfer of image information into efficient representation in computer memory are presented. These methods are utilized in conjunction with a computer-controlled microscope stage to scan the specimen field-by-field. An example of the measurements possible with this system is provided by presentation of a DNA content histogram from analysis of over 50 Because this system is capable of analyzing attached cells it offers an alternative to flow cytomeiry which requires that cells be in suspension.

This paper presents a real time image processor for obtaining threedimensional( 3-D) distance information from image sequence caused by a moving area sensor. The processor has been developed for an automated visual inspection robot system (pilot system) with an autonomous vehicle which moves around avoiding obstacles in a power plant and checks whether there are defects or abnormal phenomena such as steam leakage from valves. The processor detects the distance between objects in the input image and the area sensor deciding corresponding points(pixels) between the first input image and the last one by tracing the loci of edges through the sequence of sixteen images. The hardware which plays an important role is two kinds of boards: mapping boards which can transform X-coordinate (horizontal direction) and Y-coordinate (vertical direction) for each horizontal row of images and a regional labelling board which extracts the connected loci of edges through image sequence. This paper also shows the whole processing flow of the distance detection algorithm. Since the processor can continuously process images ( 512x512x8 [pixels*bits per frame] ) at the NTSC video rate it takes about O. 7[sec] to measure the 3D distance by sixteen input images. The error rate of the measurement is maximum 10 percent when the area sensor laterally moves the range of 20 [centimeters] and when the measured scene including complicated background is at a distance of 4 [meters] from

Transputers are processors which are well suited for the parallel implementation of medium and highlevel image processing algorithms. The serial data Links may well be used for data mterchange within the image processing task. They form however a bottleneck with respect to loading and unloading of mass data with video data rates. During the last years a parallel Transputer system was developed and implemented at the University of Technology in Hamburg-Harburg. Besides the processing power the very important aspect of communication power and topology was considered when designing the architecture of the system. As a result the system has three different communication structures: one for algorithmic data exchange one for loading and unloading of mass data like images and one for system communication of the distributed operating software. After an introduction and some architectural assumptions this paper introduces this system including its communication structures and presents the implementation of a fast algorithm for the estimation of motion parameters in image sequences.

A multispectral and low cost/performance ratio vision machine is here proposed. Integration and fusion of visual information are the main capabilities required to this architecture. Modulanty and system upgrading features are also taken in consideration in order develop a low cost/performance ratio vision machine. Two intimately linked hardware environments constitute the nodes ofthe proposed architecture: a VMEbus based subsystem is taylored for data acquisition, preprocessing, and general purpose number crunching, while an Apple MacIntosh is devoted to data and image display and to control the acquisition and processing procedure running on the VMEbus. Moreover the MacIntosh environment can be viewed as the software counterpart of the VMEbus for the modularity of its software packages. The system cost and performances are evaluated on the base of the market offering. The proposed architecture is taylored for the solution of a typical remote sensing problem: the terrain classification on the base of its thermal inertia.

A liquid-crystal optical modulator called the 2DPL-LCLM is designed to perform 2D pattern-logic operations that can be used in optical digital computations. The structure includes a layer with an array of patterned transparent electrodes between the dielectric mirror and the light-blocking layer. The utility of the device is demonstrated with an application to truth-table look-up symbolic substitution in digital optical computing that demonstrates resilience to the pattern-shift of the technique

A programmable hybrid binary image prossor based on morphological transformations is described and its applications are discussed. Using liquid crystal light modulators the fundamental operations of morphological transformations are realized optically.

In this paper a new approach to motion analysis from a sequence of stereo images has been proposed. Based on a four-frame model the new concept of unified optical flow field is presented. It is a generalization of the optical flow [1] to stereo imagery. A set of fundamental equations is established to characterize the unified optical flow field. Another set of equations is then derived from which 3-D motion can be reconstructed via the use of the field quantities of the unified optical flow field. Unlike the current existing techniques for motion analysis from stereo imagery the recovered motion by using this new approach is for a whole continuous field instead of only for some features . It does not need finding feature correspondence and therefore is much more efficient.

Surface inspection systems are being increasingly employed at high speed metal rolling mills for on-line real-lime detection of defects represented by variations from normal background. Detection algorithms typically involving ifitering and thresholding are used to segregate coil surface defects. Wear and deformation of work rolls used in rolling bring about gradual changes in the normal background. Consequently inspection systems must be capable of adapting to changing background conditions. The present paper proposes an engineering approach of threshold adjustment for background variations.

In this paper a new approach is proposed for industrial part recognition. The basic ideal of the method is that the 3D geometry parameters of parts are combined with their shape invariant features for recognition. An effective method is presented in this paper to extract the approximate 3D information of the parts from an image and a machine vision system for recognizing flat industrial parts based on the approach has been developped on PCVISION image processing system. The experimental results show that the approach works well for recognizing shape distorted flat parts.

Scanning count is an efficient way to recognize and count the number of components in a picture . After all points in one line of a target having been scanned, one of the end points in the next line of the target can be determined,then the new line can be scanned . At the sane time , the line of another side is checked to avoid any omission .Of course, this method is easy to detect that one target has been scanned over. Each point in a target is visited approximate two times . And break points can be jumped over automatically . The number and each area of targets are then calculated . Compared with the conventional algorithms, the new one is simple and fast, and is practicable in personal computers.

In this paper we describe the relation between several radio fre quency (RF) signals that are applied into the AOTF and the diffraction lights that are given off from the AOTF. These lights satisfy the color mixture principle. The experimental results are also given in this paper.

In this paper, general characteristics of <1x ia l stcreovision systems arc presented. The advantages and disadvantages of these systems arc examined. The experimental results with video-camera displacement and a video-camera with a 1.00111 system arc present ed and we conclude that it is more interesting to vary video-camera fornl length than to increase he displacement of t he video-camera w it h a fixed focal length.

This paper will describe a novel form of structured light which is currently being used as part of a research project on human speech production. A resolution of better than 1 mm has been achieved in all three directions when our technique is applied in making static measurements of dental impressions and the human face. The resulting three-dimensional data is suitable for the construction of acoustic models both mathematical and physical.

Automatic recognition of mechanical assembly drawings is very difficult problem in drawing recognition. So far many problems have not been solved. In this paper graphic algebra expressions are presented to describe part symbols and a new recognition method: serial numberleader linepart symbol correspondence is described to recognize assembly drawings. This recognition process is summarized as follows: First the assembly drawing is read by a scanner and translated into gray image data. Next serial numbers are separated from the image data. Recognized serial numbers are used to determine starting points of leader lines. A topologybased tracking algorithm is proposed to recognize a variety of leader lines find the corresponding part symbols. Finally the topological trackingmatching algorithm is built to track and recognize assembly drawings in terms of a generotortesting process of primitives. Because algebra expression of the part symbol covers lots of semantic features about the part symbol the tracking process is simultaneous with the recognition process. The above recognition method is implemented in our automatic recognition system . The experimental results show that our method is effective and isn''t affected by complex degree of assembly drawings.

Correlation filters with sharp delta-function correlation peaks (such as phase only filters and minimum average correlation energy (MACE) filters) do not easily recognize images on which they were not trained. We show that the MACE filter cannot recognize intermediate images of true class objects (e. g. aspect views or rotations midway between two training images). New Gaussian-MACE filters offer a solution to this problem.

Results of registering synthetic aperture radar (SAR) images are presented. A set of reference (tie) points are specified in a target image and an iterative search performed to obtain the corresponding tie points in a source image. Registration error is computed for each tie point by estimating its translational misalignment. The set of tie point error vectors are subsequently used to update the registration transformation. This processing sequence is repeated until convergence. The registration algorithm converges to the correct transformation in all cases tried so far. Tests on SAR imagery with known misalignment indicate that the algorithm can achieve simultaneously rms errors of 0. 01 0. 03 pixel and 0. 01 iii rotation translation and scale respectively.

For medical applications such as stereotactic surgery and radiation treatment accurate information about the size and location of the lesion is very crucial. Although magnetic resonance (MR) images does convey more clinical information than computerized tomography (CT) the geometry of MR image is distorted because of the inhomogeneous main magnetic field and nonlinear gradient fields. This paper deals with the correction of the MR image distortions by modeling the distortion function as certain high order polynomial equations. The coefficients of the polynomial equations can be determined by mapping a set of distorted reference points to its a priori locations. A special phantom is designed to correct the distortion in all three directions. To avoid roundoff error a subpixel mapping scheme is used which corrects not only the physical distortions but also the intensities of the image as well. Finally a special subharmonic condition is evaluated which when satisfied will make the use of MR images for robotic stereotactic procedures feasible.

Due to the relatively large voxel sizes of Magnetic Resonance Images (MRI) the organ boundaries represent an anatomy dependent mixture of multiple tissue types. Subsequently the image properties at the organ boundaries are highly inconsistent causing failure to produce closed organ contours using classical edge detectors. While it is widely recognized that solving of the boundary closure problems in MRI is essential for the automated 3-D volumetric reconstruction and quantification of the humananatomy only a few successful attempts have been reported in the past2''3''4''5. In this paper a new concept is presented which uses the incremental estimation of an edge by multi-pass application of a nonlinear multi-parameter edge detection operator. The operator is optimized using a quality criterion which estimates the continuity of the detected edges either directly based on a morphological prototype of the organ of interest or indirectly based on the percentage of fragmented edges found in an edge-enhanced image. Usefulness of the method is demonstrated on MR studies of the lower spine and human wrist.

The Plastic Deformation Model is introduced which simultaneously detects and preserves image data discontinuities while smoothing throughout delineated regions. This model is set into a new computational paradigm the Coupled Map Lattice approach which we introduce as a novel Parallel Distributed Processing technique. Theory and applications are presented together with an experimental study of the model.

The generalized Hough transform is an appropriate technique for pattern recognition problems involving incomplete information. Previous Hough-based approaches to the problem of 3-D object orientation determination from 2-D data have been limited in application to objects with straight line or vertex leatures generating possible correspondences between standard and transformed object views. This is because use of more general feature sets such as object contour points results in an unreasonable computational effort to solve for the transformation parameters each solution of the underdetermined equations generates a multiplicity of entries for the multidimensional arrays used to implement the transformation parameter solution space. Other methods which are not based on Hough table resolution but which exploit tensor invariance of imaged object features have proven useful for object identity and pose determination but these have required complete object feature data. A curve parameterization which exhibits tensor invariance under the affine transformation relating standard and transformed object views has been employed by the authors to generate local feature data for objects with partially extractable planar features. The transformation parameters are obtained from the data by solving a completely specified set oflinear equations greatly reducing the computational complexity of the Hough approach. The solutions can then be subjected to flexible constraints consistent with rigid 3-D motion orthographic projection and geometric invariance. Each tentative match generates a single entry in the Hough parameter space. The sparsely filled solution

Although physician observation is usually the most sensitive method for diagnosing and monitoring a patient''s medical condition human observation cannot be conducted continuously and consistently. It can be helpful therefore to employ specialized automated techniques for the continuous reliable and noninvasive monitoring of those parameters useful for the enhancement of physicians'' diagnostic capabilities. Signal processing systems are among the most powerful of those techniques currently available for noninvasively examining the internal structure of living biological systems. Nonetheless the capability of these systems can be substantially enhanced if supplemented with automated classification and interpretation precedures. An intelligent EEG signal sensing and interpretation system using typical signal processing techniques supplemented with heuristics and identification techniques has been designed. The system is comprised of five major components namely: the fact gathering system the knowledge/rule base the knowledge organization/learning phase the inference engine and the expert/user interface. The fact gathering system collects raw waveforms preprocesses these for noise elimination and extracts the pertinent information from the waveforms. The knowledge/rule base is an information and knowledge bank wherein the appropriate knowledge parameters useful for the decision making process are stored. The knowledge organization/learning phase structures the knowledge In the order determined by the association among pattern classes and trains the Inference engine. The structure of the inference engine is based on a hierarchical pattern classifier which categorizes the unknown signals using a layered decision making strategy

An innovative algorithm is presented for the automatic detection of spot targets in infrared terrain image sequences. The algorithm is designed to successfully perform target detection also in presence of deceitful hot background structures such as rocks and sand representing the hottest regions of the scene. The original contribution is represented by the thermal model adopted for the infrared scene and which drives the intraframe processing by iteratively applying three processing steps: i) thresholding ii) clustering iii) semantic labeling. The aim of the intraframe segmentation module is first to detect well contrasted targets and then verify the presence of less contrasted ones on the hypothesis that they all belong to the same formation. At each iteration within the intraframe analysis the thermal threshold is lowered and the hot part of the scene is processed to look for the presence of targets. By theans of different clustering procedures the hot part of the scene is partitioned into a set of connected objects each characterized by a feature vector consisting of area peak thermal value and centroid coordinates. The objects which correspond to the target model are marked as potential targets. If they have been already detected during the previous iterations without sensible changes in the descriptive parameters they are labeled as " confirmed" targets. If on the contrary the area has significantly increased they are labeled as " expanded" targets. If a

We consider the problem of point source target detection. It is desirable to remove sensor degradation from images to obtain a better object representation. Two techniques are considered: a simple four-pixel average and a more complex maximum entropy deconvolution approach. Each of the methods perform well under certain varying conditions. Our motivation is based on applications such as: (1) weapon systems for the Strategic Defense Initiative that require a capability to engage a multiple target threat at long ranges and (2) star tracking systems for satellite guidance.

The objective of this work is to develop automated techniques for recognizing the same objects in images that differ in scale tilt and rotation. Such perspective transformations of images are produced when aerial images of the same scene are taken from different vantage points. In previously reported work we have identified methods for deriving algebraic projective invariants under central projections. These methods generalize the familiar cross-ratio theorems for single images of finite sets of points on the line and in the plane. The algebraic methods do not utilize the intensity values of the images. Since image features essential for object recognition may be described in terms of derivatives and integrals of the image intensity it is necessary to investigate whether certain differential and integral operators applied to different perspective views of the same object are also invariant under the perspective transformation. We proceed to derive new differential operators and their corresponding integral invariants for curves and planar objects. Extensions to other image formation models such as synthetic aperture radar (SAR) are discussed. These results are steps toward a computational model for perspective-independent object recognition.

Correlation scene matchers used for image registration are subject to geometric distortions of the sensed field-of-view (FOV) and reference images. These distortions limit the size of the optimal FOV. Since larger FOVs mean more information and thus higher correlation peak-to-sidelobe ratios (PSR) we have been motivated to find a technique for further increasing effective FOV size. We have developed a method of forming mosaic FOVs for individual FOV in concert with short-term stable navigation data that specify the relative displacements of the individual FOVs. This technique called correlation addition is useful in scene matching navigation where a camera is scanned (e. g. by platform motion) over a region of interest. Correlation addition can be mechanized by translation and addition of the correlation surfaces from individual FOV correlations. We have implemented a VLSI architecture for computing correlation addition and have conducted performance simulation of scene-matching navigation in stressful conditions such as scene changes due to snow and shadows. We will present performance analysis several algorithms for correlation addition some simulation results and describe our architecture for computing the correlations.

A generalized quadratic (Bayesian-like) classification system has been developed for evaluating the performance of other classifiers such as neural networks in automatic target recognition (ATR). The system was tested using multispectral real data as well as computer generated data sets. The classifier employs the covariance matrix and centroid of the feature set to describe each region. The system then calculates the likelihood associated with an unknown object belonging to a defined region. A multivariate normal distribution is assumed in calculating this likelihood. The system utilizes a learning algorithm to continuously upgrade performance and has shown near 100 percent accuracy even after very short training periods.

In moire methods and other full field measurement techniques suitable for robot vision it is often of key importance to be able to distinguish rigid body transformations from local deformations or defects. This is especially the case when using a computer generated surface as the reference(master). One solution is to maneuver the test object in space until the rigid body displacements are eliminated. This is however a very time consuming process. In this paper an algorithm is presented for automatic transformation in space ( 6 degrees of freedom ) of the computer generated surface. The aim is to position the master object in space such that maximum overlap between the master and the test object is achieved. Basically this system consists of an illumination unit a CCD camera a digital image processor and computer-generated master objects.

Cornea! endothelium cells of cats eyes have been observed and analysed by means of a PC based image processing system. The dimensions such as the surface S and the perimeter P as well as the density D and deformations describted by two factors F1 I and F2 4 it S I P2 have been statistically evaluated for two classes of 120 cell photographs corresponding to two different treatements A and B . Similar evaluation can be extended to humain eye comeal cells. The aim of this communication is to present a new algorithm for calculating the surface and perimeter of cells. The method is based on encoding the geodesic centre of the cells in terms of mathematical morphology with different grey levels which is then propagated onto the entire cell up to the cell edge. The function of the histogram is then calculated in order to determine the cell parameters. The advantage of this method is rapidity and no calculating error for the binary cell image. This algorithm of automatic diagnostic for endothelium cell analysis has been tested by means of a reference image. The results are represented by a series of photographs.

A recently introduced tree growth algorithm the Marginal Returns (MR) algorithm is used to grow multiple rate tree structured vector quantizers for the pyramid coding of hexagonally sampled images. The use of a structured multi-rate code book solves two problems that normally arise in vector quantization of subbands. The multiple rate code book can operate over a wide range of rates thus dispensing with the need to transmit the code book as overhead while the tree structure reduces the search complexity. Search complexity is a crucial issue even in low rate pyramid coding since subbands with more information content are coded at high rates. In addition the design technique makes it possible to tune the coder to the spectral properties of the image by optimally allocating rate to the different subbands. It has been shown in an earlier paper that the Marginal Returns algorithm yields code books that are optimal for sources that meet the law of diminishing marginal returns. However even for sources that do not satisfy these conditions the algorithm gives coders that perform close to the optimal. Image coding results at rates below 1 bpp are presented.