White point mapping color correction is compared to a method which performs color correction using knowledge of the spectral reflectance distribution in the image. The methods are compared using chromaticity plots and the CIE L*a*b* (Delta) E measurement.

This paper illustrates the use of Hopfield neural networks to halftone color images. We define an error function which is the weighted sum of squared errors of the Fourier components of the original and halftoned images. The weights can be chosen to match the human visual system or other input/output transfer functions. The error function is minimized by using a neural network and solving its dynamical equation iteratively. FFTs are used to perform the necessary convolutions so that the computational requirements are reasonable even for large images.

We propose an arithmetic coding model for color images processed by the error diffusion. The error diffusion is a dynamic thresholding method of high quality tone rendering. However, there is a problem that probability model of conventional compression methods such as MH and MMR are not suitable for bi-level images processed by the error diffusion. Consequently, much transmission time or much accumulation memory is required. We have constructed a probability model which is based on the processing structure of the error diffusion, and combined arithmetic coding because of its flexibility in the construction of probability models. Simulation results have shown that this coding model is efficient for color images processed by the error diffusion.

This paper presents a feature of an image data compression system, and the algorithm employed in a newly developed digital still camera (FUJIX Card Camera DS-H1). DS-H1 is composed of a CCD with 390K pixels, a digital signal processing unit, and image data compression units named DCT and CODER. An IC memory card of 8 Mbit SRAM is used as the recording media. DS-H1 has not only a function of an image data recording, but also a play back for the instantaneous review. A new algorithm is used for controlling a bit rate of a still frame compressed by ADCT (Adaptive Discrete Cosine Transform). The image data control is an important function for digital still camera in order to guarantee the total number of still images stored in an IC memory card. An information distribution of each image data blocks (8*8) is estimated as ACTIVITY. An adequate quantization step ((alpha) ) is calculated using ACTIVITY and parameters named K1, K2. A set of K1 and K2 suitable for the target compression ratio is transferred from CPU to CODER before the compression sequence starts. As a result of our evaluation, it is actually proved that ADCT with the control function is successfully implemented in a small camera body of DS-H1. Picture quality of the played back images is good. K1, K2 and ACTIVITY are important parameters to determine the bit rate control accuracy and the picture quality of compressed/decompressed images. Our simulation results shows that adequate K1, K2 realize a good picture quality and a high control accuracy.

An adaptive motion-compensated spatial-temporal filtering technique is presented. It is shown as an effective control parameter to gracefully tradeoff the quality of DCT coded video and bit rate in applications such as the MPEG video coding standard from digital storage media. Conventionally, spatial and temporal filters are used as fixed prefiltering parameters to reduce image resolution for coding and to eliminate background noise. This paper presents a new method to dynamically adjust the filters from the rate-control buffer. The filtering is performed over the motion path. Spatial filtering is also performed to provide smoother motion flow. Thus, a tradeoff of image quality and bit rate can be better balanced. By applying this feed-back control, the maximum image resolution is maintained for normal conditions while a graceful degradation is achieved for difficult scenes. Simulation results based on the MPEG video algorithm operating at .55 Mbit/sec and .95 Mbit/sec show that the proposed scheme is effective in alleviating block artifacts.

We propose a new digital halftoning algorithm where the halftoning is achieved by a pixelwise comparison of the gray scale image against a nonimage array, the blue noise mask. The blue noise mask is constructed such that when thresholded at any level, the resulting binary pattern has the correct first order statistics, and also its power spectrum has blue noise (high frequency) characteristics which are visually pleasing. The construction of the blue noise mask is described and experimental results are shown. Also results from a phychovisual study are provided where subjects rated halftoned images that have the same first order but different second order statistics.

Recently, a very simple nonlinear algorithm, the so-called Teager's algorithm, has been introduced to calculate the energy of a 1D sequence. In this paper, this algorithm is extended to 2D signals and applied to the constant enhancement of images. Initial simulation results indicate that the proposed approach yield a superior and visually pleasing enhancement of natural images. The new algorithm is very simple and permits fast implementation.

The class of Intensity-Dependent Spread (IDS) filters proposed by Cornsweet and Yellott produces the reflectance ratio map at the output of the filter independent of illumination. We propose a new method to optimize the spread function in the context of the IDS filters. The optimum spread is the solution to a variational formulation where the image noise is minimized subject to a smoothness constraint. In our solution, the Lagrangian parameter is space-dependent and also is inversely proportional to the spread function. This solution is a function of the input image and its first and second derivatives. The optimized scale-function is then applied to the IDS filter structure to produce sharp edge localization as well as reflectance ratio estimates independent of illumination. The simulation results illustrate the fact that the Optimum Intensity-Dependent Spread filter improves the performance of the IDS filter and also is two orders of magnitude faster for 512 X 512 images. Examples comparing the results of the two filter structures are illustrated.

An efficient transformation algorithm for 3D images is presented. The algorithm is fairly amenable to parallelization. Its implementation is a transputer network will be discussed. The presented algorithms are necessary in many application areas, such as medical imaging and landscape imaging.

A model-based Kalman estimator is developed for spatial-temporal filtering of noise and other degradations in velocity and depth maps derived from image sequences or cinema. As an illustration of the proposed procedures, edge information from image sequences of rigid objects is used in the processing of the velocity maps by selecting from a series of models for directional adaptive filtering. Adaptive filtering then allows for noise reduction while preserving sharpness in the velocity maps. Results from several synthetic and real image sequences are given.

In this paper, the applications of the iterative Gauss-Newton (GN) approach in nonlinear image restoration are considered. The convergence properties of a general class of nonlinear iterative algorithm are studied through the Global Convergence Theorem (GCT). The iterative GN algorithm for the solution of the least-squares optimization problem is presented. The computational complexity of this algorithm is enormous, making its implementation very difficult in practical applications. Structural modifications are introduced, which drastically reduce the computational complexity while preserving the convergence rate of the GN algorithm. With the structural modifications, the GN algorithm becomes particularly useful in nonlinear optimization problems. The convergence properties of the algorithms introduced are readily derived, on the basis of the generalized analysis and the GCT. The applications of these algorithms on practical problems, is demonstrated through an example.

Attention is given to a SAR wave equation-based system model that accurately represents the interaction of the impinging radar signal with the target to be imaged. The model is used to estimate the complex phase error across the synthesized aperture from the measured corrupted SAR data by combining the two wave equation models governing the collected SAR data at two temporal frequencies of the radar signal. The SAR system model shows that the motion of an object in a static scene results in coupled Doppler shifts in both the temporal frequency domain and the spatial frequency domain of the synthetic aperture. The velocity of the moving object is estimated through these two Doppler shifts. It is shown that once the dynamic target's velocity is known, its reconstruction can be formulated via a squint-mode SAR geometry with parameters that depend upon the dynamic target's velocity.

We propose a new iterative block reduction technique based on the theory of projection onto convex sets. The basic idea behind this technique is to impose a number of constraints on the coded image in such a way as to restore it to its original artifact-free form. One such constraint can be derived by exploiting the fact that the transform coded image suffering from blocking effects contains high frequency vertical and horizontal artifacts corresponding to vertical and horizontal discontinuities across boundaries of neighboring blocks. Since these components are missing in the original uncoded image, or at least can be guaranteed to be missing from the original image prior to coding, one step of our iterative procedure consists of projecting the coded image onto the set of signals which are bandlimited in the horizontal or vertical directions. Another constraint we have chosen in the restoration process has to do with the quantization intervals of the transform coefficients. Specifically, the decision levels associated with transform coefficient quantizers can be used as lower and upper bounds on transform coefficients, which in turn define boundaries of the convex set for projection. Thus, in projecting the 'out of bound' transform coefficient onto this convex set, we will choose the upper (lower) bound of the quantization interval if its value is greater (less) than the upper (lower) bound. We present a few examples of our proposed approach.

Image restoration results that are both objectively and subjectively superior can be obtained by allowing the regularization to be spatially variant. Space-variant regularization can be accomplished through iterative restoration techniques. The optimal choice of the regularization parameter is usually unknown a priori. The generalized cross-validation (GCV) criterion has proven to perform well as an estimator of this parameter in a space-invariant setting. However, the GCV criterion is prohibitive to compute for space-variant regularization. In this work, we introduce an estimator of the GCV criterion that can be used to estimate the optimal regularization parameter. The estimator of the GCV measure can be evaluated with a computational effort on the same order as that required to restore the image. Results are presented which show that this estimate works well for space-variant regularization.

In many modern applications, the image data consists of a sequence of frames which may be degraded by object motion. Examples include high- speed television where the purpose is to record and analyze rapid motion of objects and production of still frame hardcopy from vide recordings. It is often the case that the motion of the image of the object on the recording sensor moves significantly during the time the shutter is open. Since different objects may be moving in different directions at different speeds, the resulting blur is space-variant. Two problems are addressed in this research: determination of the space-variant motion and restoration with space-variant point spread functions. The work presented here uses the surrounding frames to obtain information about the relative motion, and together with information about the shutter speed, estimates the motion blur in local regions of the image. The advantage is that there are no limitations of the type of motion which can be treated. A restoration method has been developed which can deal with this directly, as opposed to geometrically warping the image to produce an image which is treated by spatially-invariant methods. The method is a modification of the iterative Landweber iteration which is implemented using overlapping sections.

In this paper, we address the identification of a class of continuously varying linear space variant (LSV) blurs, for which the variation of the PSF parameters can be tracked from the loci of the zero-crossings in the magnitude of the local Fourier transform (LFT) of the blurred image. We show that this class of LSV blurs can be identified by an LFT analysis provided that the variation of the PSF parameters over the image is sufficiently slow. It is demonstrated that the size of the LFT analysis window needs to be approximately four times the blur support to obtain good results. However, since the blur size is unknown to start with, we propose a multi-(spatial) resolution LFT analysis of the blurred image to determine the best window size and the estimate of the blur support. We provide experimental results for the case of LSV motion blur.

A Bayesian approach to image reconstruction from emission tomography image is presented in which the image is modeled using a joint Gibbs distribution of emission intensities and line processes. The line process represents the presence or absence of discontinuities between each neighboring pair of pixels. It is introduced to avoid the smoothing across discontinuities, which commonly occurs in Bayesian image estimation when a line process is not included. Two algorithms for MAP estimation over both intensity and line processes are presented. Both methods employ the generalized EM (GEM) algorithm to avoid direct optimization over the posterior distribution which does not share the Markovian property of the prior. The M-step of the GEM algorithm of the MAP estimation problem requires optimization over a function which has the appealing property that the neighborhood is identical to that of the prior. During the M-step both the intensity and line processes are updated. This is achieved in two stages. In the M1-step the intensities are updated, while holding the line process constant, using a gradient descent method. In the M2-step the line process is updated, with the intensity process held constant. Two alternative M2-steps are described in the paper. The use of a line process in the image model also provides a natural framework for the incorporation of a priori information from other modalities. In this case, boundaries may be found from MR or CT images and used as known line processes in the image estimation procedure.

Algorithms that are used in image reconstruction fall broadly into one of two categories [1-2]. First category is based on Fourier methods to reconstruct a function from its line integrals on certain regions. Second category is based on probabilistic reconstruction methods which take into account the stochastic nature of the problem and hence yield more accurate results under a broad range of parameters and data. In most general terms image reconstruction methods take the data that is passed through a general filtering mechanism and reconstruct the original image. In medical applications the amount of data that should be processed for this purpose is usually prohibitive for certain methods due to storage and time limitations. In this paper we will describe methods to help overcome these difficulties using various techniques. Positron emission tomography (PET) is a medical diagnostic procedure [3-4] that enables physicians to visually evaluate the metabolic activity in various organs. Rather than generating 'static' pictures as in X-rays, PET introduces low levels of positron emitting radioactive material in the organ under study, and levels of absorption on various parts of the organ is measured by the PET scanner [5]. The type of biochemical and the radioactive material used depends on the organ to be studied. It is known that the brain uses glucose as a primary energy source and therefore glucose 'injected' with radioactive material is used for brain studies, for the study of heart deoxyglucose and palmitic acid injected with the radioactive material is used. Depending on the met.abolization of the injected material one can generate a 'dynamic' picture of the organ. For example if the brain is studied, the patient's psychosomatic condition at the time of the study will be different than at a different time and condition. This hopefully will reveal valuable information on the patient's condition and the effect of various treatments (schizophrenia, etc.). These is a broad literature on PET imaging and the interested reader is referred to [4-5] and the references there.

Suppose that it is desired to estimate certain parameters associated with a model of an object that is contained within a larger scene and that only indirect measurements of the scene are available. The optimal solution is provided by a Bayesian approach, which is founded on the posterior probability density distribution. The complete Bayesian procedure requires an integration of the posterior probability over all possible values of the image exterior to the local region being analyzed. In the presented work, the full treatment is approximated by simultaneously estimating the reconstruction outside the local region and the parameters of the model within the local region that maximize the posterior probability. A Monte Carlo procedure is employed to evaluate the usefulness of the technique in a signal-known-exactly detection task in a noisy four-view tomographic reconstruction situation.

A new iterative algorithm for restoring noisy blurred images with unknown point spread function (PSF) is presented. The method initially estimates the PSF and the original image with the Expectation Maximization (EM) algorithm. The resulting image estimate is then refined by using the adaptive Row Action Projection (RAP) algorithm which is based on the theory of Projection Onto Convex Sets (POCS). The new implementation of the RAP can be performed efficiently in parallel and facilitates locally adaptive constraints and cycling strategies. Computer simulations illustrate the new method to be very competitive in restoring degrading images from noisy blurred images with unknown PSF.

Finite-dimensional linear parametric models for multidimensional random signals have been found useful in many applications such as image coding, enhancement, restoration, synthesis, classification,a nd spectral estimation. A vast majority of this work is based upon exploitation of only the second-order statistics of the data either explicitly or implicitly. A consequence of this is that either the underlying models should be quarter-plane (or, half plane) causal and minimum phase, or the impulse response of the underlying parametric model must possess certain symmetry (such as 'symmetric noncausality'), in order to achieve parameter identifiability. I consider a general (possibly asymmetric noncausal and/or nonminimum phase) 2D autoregressive moving average random field model driven by an independent and identically distributed 2D non-Gaussian sequence. Several novel performance criteria are proposed and analyzed for parameter estimation of the system parameters given only the output measurements (image pixels). The proposed criteria exploit the higher order cumulant statistics of the data and are sensitive to the magnitude as well as phase of the underlying stochastic image model.

We described a method for automatically identifying and separating pixels representing bone from those representing soft tissue in a dual- energy point-scanned projection radiograph of the abdomen. In order to achieve stable quantitative measurement of projected bone mineral density, a calibration using sample bone in regions containing only soft tissue must be performed. In addition, the projected area of bone must be measured. We show that, using an image with a realistically low noise, the histogram of pixel values exhibits a well-defined peak corresponding to the soft tissue region. A threshold at a fixed multiple of the calibration segment value readily separates bone from soft tissue in a wide variety of patient studies. Our technique, which is employed in the Hologic QDR-1000 Bone Densitometer, is rapid, robust, and significantly simpler than a conventional artificial intelligence approach using edge-detection to define objects and expert systems to recognize them.

An algorithm is described which segments magnetic resonance images while removing the noise from the images without blurring or other distortion of edges. The problem of segmentation and noise removal is posed as a restoration of an uncorrupted image, given additive white Gaussian noise and a segmentation cost. The problem is solved using a strategy called Mean Field Annealing. An a priori statistical model of the image, which includes the region classification, is chosen which drives the minimization toward solutions which are locally homogeneous and globally classified. Application of the algorithm to brain and knee images is presented.

A new approach is developed for detection of image objects and their orientations, based on distance transforms of intermediate level edge information(i.e., edge segments and vertices). Objects are modeled with edge segments and these edges are matched to edges extracted from an image by correlating spatially transformed versions of one with a distance transform of the other. Upper bounds on change in cross- correlation between edge maps and distance transforms are shown to be simple functions of change in translation and rotation. The process of computing the optimal object rotation at each possible translation can be accelerated by one to two orders of magnitude when these bounds are applied in conjunction with an object translation-rotation traversal strategy. Examples with detection and acceleration results demonstrate the robustness and high discriminatory power of the algorithm.

The aim of this paper is to find a relationship between alternating sequential filters and the morphological sampling theorem developed by Haralick. First, we show an alternative proof for opening and closing in the sampled and unsampled domain. This is done by using basis functions. This decomposition is used then to show the relationship of opening- closing in the sampled and unsampled domain. An upper and a lower bound, for the previous relationships, were found. Under certain circumstances, an equivalence is shown for opening-closing between the sampled and the unsampled domain. An extension to more complicated algorithms is also considered, namely; union of openings and intersection of closings. The reason to consider such transformations is that in some applications we would like to eliminate pixels removed by individual openings (closings).

Kalman filtering is a methodology that has demonstrated great potential for solving a large number of problems in many areas. It has so far had its greatest success in the areas of control theory and process control, but is a method and it is not limited to this area. There are a great number of areas within oil and gas exploration where it can prove to be of great success. Of particular interest is an aid to exploration in mature hydrocarbon provinces. The identification of remaining reserves of hydrocarbons in stratigraphic traps in the world's mature hydrocarbon provinces is a difficult task. Often these traps are small compounding identification. A subset of this general problem is the appropriate location of new wells in known producing areas. The task reported on herein has been to complete an uncompleted pattern of successful and unsuccessful wells. Two dimensional Kalman filter and interpolation theory is used to estimate successful and unsuccessful well locations. Based on a map of the known well locations, the image representing the estimated pattern is completed. The majority of image pixels in this particular case will be unknown, and not just distorted off of its original value by noise. Examples are detailed and discussed.

Determination of an optimal window to improve the performance for registering nonlinear distorted images using a cross-correlation technique is presented in this paper. A 2D cross-correlation technique is applied to two meteorological radar images (the search and reference images) which possess the characteristics of nonlinearity, geometric distortion, and ever-evolving pattern. Various sizes of concentric square windows of the reference image are used for computing the cross- correlation field. Parameters of cross-correlation field such as peak value, location of the peak, and standard deviation are determined. The location of the peak correlation, instead of its peak value, is chosen as an indicator which best describes the performance of registration, since the location remains unchanged for certain sizes of the window. This location represents the translation shift of the images or the offset of the registration. These windows cover a major portion of the autocorrelation area of the reference image. The standard deviation of the cross-correlation field between search and references images its maximum for these window sizes which are considered as optimal.

In this paper, a method is developed to performing the textural analysis and use of contextual information in classification of satellite imagery, with SPOT images. As a first step, an efficient method is used to determine for each region about twenty statistical parameters, using for some of them fractal-based techniques. As a second step, controlled segmentation is performed as follows: Use of high level knowledge to improve previous segmentation, with squared unit regions of 16 X 16 pixels. At this stage different kinds of knowledge can be used: Contextual decision rules. Cartographic or climatological information.

Early vision processing involves segmenting and eliminating/minimizing extrinsic variations in an image. Tradition segmentation algorithms use a single threshold to reduce image data to the regions important to the vision system. A single threshold rarely gives good segmentation results over an entire image because it may reject too many of the objects as belonging to the background or may accept too much of the background as belonging to the objects. This paper presents a segmentation method that uses multiple thresholds for partitioning an image into objects and background. The segmentation process is a gradient approach whereby the multiple (or adaptable) thresholds are calculated from discriminant structural feature values.

This paper proposes a new algorithm for detecting character strings in an image containing illustrations and characters. It also describes a part number entry system that utilizes this algorithm. The algorithm detects character strings by investigating the horizontal boundaries of rectangles representing characters strings. It can be performed a high speed, and can detect characters touching an illustration. Using this algorithm, the part number entry system extracts areas of part numbers scattered among illustrations and then recognizes the. This is a software program implemented on a personal computer, and is composed of four subprograms: detection of character strings, character recognition, post-processing, and flexible user-interface for error correction.

This paper describes a PC-based near-real time implementation of a two- channel maximum likelihood classifier. A prototype for the detection of ice formation on the External Tank (ET) of the Space Shuttle is being developed for NASA Science and Technology Laboratory by Lockheed Engineering and Sciences Company at Stennis Space Center, MS. Various studies have been conducted to obtain regions in the mid-infrared and the infrared part of the electromagnetic spectrum that show a difference in the reflectance characteristics of the ET surface when it is covered with ice, frost or water. These studies resulted in the selection of two channels of the spectrum for differentiating between various phases of water using imagery data. The objective is to be able to do an online classification of the ET images into distinct regions denoting ice, frost, wet or dry areas. The images are acquired with an infrared camera and digitized before being processed by a computer to yield a fast color-coded pattern, with each color representing a region. A two- monitor PC-based setup is used for image processing. Various techniques for classification, both supervised and unsupervised, are being investigated for developing a methodology. This paper discusses the implementation of a supervised classification technique. The statistical distribution of the reflectance characteristics of ice, frost, water formation on Spray-on-Foam-Insulation (SOFI), that covers the ET surface, are acquired. These statistics are later used for classification. The computer can be set in either a training mode or classifying mode. In the training mode, it learns the statistics of the various classes. In the classifying mode, it produced a color-coded image denoting the respective categories of classification. The results of the classifier are memory-mapped for efficiency. The speed of the classification process is only limited by the speed of the digital frame grabber and the software that interfaces the frame grabber to the monitor. The process has been observed to take 4 seconds for a 512 X 480 pixel image. This set-up may have applications in other areas where detection of ice and frost on surfaces is of critical importance.

In a digital video tape recorder (DVTR), the video data are stored serially on a sequence of parallel scan tracks running obliquely across the tape. In normal playback, the read heads scan along each track sequentially, while fast-forward is realized by a special subsampling process in which the read heads sweep across multiple tracks per scan and pick up a segment of data from each track to generate a patchwork quilt of subimages taken from successive fields. An image compression algorithm for DVTR is required to conform to the storage format implied by this fast-forward mechanism. Thus, a field of video is partitioned into a small number of subimages, each independently coded with a fixed number of bits. This requirement excludes the use of interframe coding and most variable-rate coding algorithms. We propose a new image coding algorithm that retains the efficiency of variable-rate coding while the above requirement is satisfied. In this algorithm a subimage is partitioned into non-overlapping 8 X 8 blocks, and each block is coded progressively by DCT and Vector Quantization to generate a finite set of candidate output blocks covering a range of bit rates. A procedure based on the Lagrange Multiplier method is used to select one of the candidate outputs for each 8 X 8 block such that the overall distortion is minimized subject to the constraint on the bit rate. Simulation results demonstrate the potential of this algorithm in image compression for DVTRs.

We have undertaken a study of techniques for the perceptually transparent coding of very high quality and resolution images. These techniques should be free of any of the visual artifacts due to transform coders, DPCM coders or vector quantization, so that post processing to improve image quality or resolution can be performed. The approach starts with a decomposition of an image into a spline approximation, based on a subsampled array of pixels. The spline approximation is then subtracted from the original and the resulting remainder is quantized non-uniformly and with as few levels as possible while still insuring perceptual transparency. This differential quantization takes advantage of properties of human vision. The coarsely quantized remainder is then encoded in an error free fashion. Several techniques are being considered for this error free coding. Since the only errors are introduced in the quantization of the remainder, the errors are not perceptible and there is no structure to them.

This paper deals with both the theoretical and the practical aspects of information preserving image data compression. The concept of entropy and the common effective compression algorithms are first reviewed. The first-order, second-order and the conditional entropies of images are calculated and compared to the compression level of some existing algorithms. A set of reversible transformation techniques are then introduced and applied to the referenced images. The purpose of these transformations is to lower the entropy in order to improve the compression ratio of the compression algorithms. The effect of the transformations on the compression algorithms are examined. Finally, suggestions for further research in lossless compression are given.

This paper reports on a current investigation of a new approach to vector quantization (VQ) for image coding. The approach consists of integrating finite state vector quantization (FSVQ) into a binary residual vector quantizer structure. Due to the inherently predictive nature of FSVQ, the quality of image representation is often greatly superior to that of memoryless VQ (for small vector sizes). However, the codebook storage requirements are typically very large by comparison. The objective of this work is to reduce the codebook storage requirement of FSVQ without significantly impairing the reproduction quality of the coded image. Experimental results indicate that low bit rate systems based on a combination of FSVQ and RVQ can be designed with performance approaching the quality of the FSVQ schemes but with only a very small fraction of the storage requirement.

An important part of most video compression techniques is displacement vector estimation, which in many situations is achieved by Block Matching Algorithms (BMA). Traditional BMA uses costly integer arithmetic on 8-bit images to minimize a Mean Absolute Error (MAE) criterion that determines the displacement estimate. In this paper, we present an approach for motion estimation based on binary edge images. Specifically, we assign an intensity gradient vector to each edge pixel and apply BMA to the gradient vector. Our approach can potentially reduce the complexity of the VLSI implementation of motion estimation algorithms. We present simulation results comparing this technique to traditional BMA over several 50-frame sequences.

Visual Pattern Image Sequence Coding (VPISC) is a pyramidal image coding scheme which utilizes human visual system (HVS) properties to achieve low bit rates while maintaining a good perceived image quality, all with extremely low computational cost. This paper describes extensions of VPISC, termed Foveal VPISC (FVPISC) and Adaptive VPISC (AVPISC). Both algorithms produce decreased bit rates by selectively allowing some image regions to be coded at low resolution. In FVPISC, a foveation criterion is used to select a region of interest. In AVPISC, the algorithm adaptively determines which regions require high-resolution coding in order to maintain uniform image quality over the entire image. After coding but before transmission, the encoder examines its own output, and reduced the transmission bit rate by eliminating high- resolution information about low-resolution portions of the image. The method is adaptive in the sense that the bit rate is locally adjusted to match the level of detail present in an image region.

Hilbert scanning defines a mapping, hn : R yields Un, that maps the unit interval onto the n-dimensional unit hypercube continuously. In the discrete case the mapping can be described in terms of Reflected Binary Gray Codes (RBGC). In order to extend the quantized mapping to arbitrary precision it is necessary to define induction rules. Induction rules are defined in terms of a single canonical sequence and a set of rotations. In general, in an n-dimensional hypercube there are n2n possible orientations of a canonical form. Beyond two dimensions, it is possible to have nontrivially different paths between two possible orientations and it is better to define the induction rule in terms of the end points of the RBGC subsequences. Hilbert coding is used for n- dimensional binary data compression. The effectiveness of this method to data compression is confirmed. Experimental evaluation shows Hilbert- Wyle coding to be consistently better than other standard compression methods.

A video coding technique aimed at compression of video and its associated audio at about 1.5 Mbits/s has been developed by the Moving Pictures Expert Group (MPEG-Video). The Moving Pictures Expert Group is part of the ISO-IEC JTC1/SC2/WG11, an organization responsible for the standardization of coded representation of video and audio for information systems. The video compression technique developed by MPEG covers many applications from interactive systems on CD-ROM to delivery of video information over telecommunications networks. The MPEG video compression algorithm relies on two basic techniques: block based motion compensation for the reduction of the temporal redundancy and transform domain based compression for the reduction of spatial redundancy. Motion compensation techniques are applied with both predictive and interpolative techniques. The prediction error signal is further compressed with spatial redundancy reduction (DCT). The quality of the compressed video with the MPEG algorithm at about 1.5 Mbits/s has been compared to that of consumer grade VCR's. The quality, cost and features of the MPEG video algorithm make it directly applicable to personal computers and workstation thus allowing the development of many new applications integrating video, sound, images, text and graphics.

We introduce a new scaled Discrete Cosine Transform (SDCT) and inverse SDCTs optimized for architectures where a primitive arithmetic operation is a fused multiply/add. Explicit algorithms are derived for 1- dimensional inputs of 8 points and for 2-dimensional inputs of 8 X 8 points. The latter require 416 operations. When constants are programmable, descaling plus computing the inverse DCT on 8 X 8 points can be done with 417 operations.

Most image coding algorithms, like the P X 64 and MPEG-1 standards, use locally derived estimates of object motion to form a prediction of the current frame. But camera motion, such as zooms and pans, which systemically affect the entire frame, is seldom handled efficiently. In this paper, we study the modeling, estimation and compensation of global motion caused by camera zooms and pans, we model the global motion in each frame with just two parameters: a scalar zoom factor and a 2D pan vector. Parameter estimation minimizes the squared prediction error of either the difference frame or the optical flow field. The estimated parameters are then used to construct a zoom/pan compensated prediction of the current frame, upon which some local motion compensation algorithm can then be applied to model object motion. Simulations suggest that these two global motion estimation algorithms are robust and accurate, and that global motion compensation provides a better prediction of the current frame with a potentially large reduction of motion side information.

This paper describes a flexible parallel processing architecture designed for use in real time video processing. The system consists of floating point DSP processors connected to each other via fast serial links, each processor has access to a globally shared memory. A multiple bus architecture in combination with a dual ported memory allows communication with a host control processor. The system has been applied to prototyping of video compression and decompression algorithms. The decomposition of transform based algorithms for decompression into a form suitable for parallel processing is described. A technique for automatic load balancing among the processors is developed and discussed, results ar presented with image statistics and data rates. Finally techniques for accelerating the system throughput are analyzed and results from the application of one such modification described.

Conventional subband coding for image data compression uses 2D separable QMF banks in which the analysis and synthesis filters are composed of 1D filters. Such an implementation produces a large size output image as a result of the convolution process. Various signal extension methods have been proposed to solve this problem. However, these methods have one or more of the following drawbacks: generation of boundary noise, inability to guarantee aliasing cancellation, and increased computation complexity. In this paper, we present an alternative solution to the problem by converting a 2D image array to a 1D array and then using a 1D QMF bank to process the 1D signal. In our approach, most of the above drawbacks mentioned above are eliminated. In addition, our approach offers more flexibility in the type of the filter that can be implemented.

Two channel conjugate VQ (TCCVQ) has been recently proposed for speech coding. Compared to the ordinary VQ, two channel conjugate VQ requires less computational complexity, less storage (memory) requirement and is more robust against channel errors [8], [16], [17]. This technique is now extended to image coding. Initial investigation confirms these advantages of two channel conjugate VQ. For noise free channel the performances of both techniques are similar. However, when channel noise is injected, for the same bit rate, images reconstructed based on two channel conjugate VQ are subjectively more pleasing (less visible distortion) compared to those based on ordinary VQ. Both VQ techniques are also applied to a well known separating mean VQ (SMVQ) [2], [20] and are tested for robustness against channel errors. Computer simulations demonstrate the success of the two channel conjugate VQ technique for image data compression at various bit rates.

This work develops iterative algorithms for decoding cascade-coded images by Relative Entropy (RE) minimization. In cascade coding, blocks of an image ar first transform-coded and then the retained coefficients are transmitted by using moment-preserving Block Truncation Coding (BTC). The BTC coding introduces a quantization error in the values of the retained coefficients. Upon reception,t he distorted coefficients are used in reconstructing the image by the inverse transform, with the unretained coefficients set equal to zero. The proposed algorithms construct the original image from the distorted coefficients by minimizing the RE of the image, with the coefficients used as constraints. In addition, the error introduced by the BTC coding is used as an additional constraint, since it is known to the receiver by the nature of the BTC coding. The iterative nature of the algorithm pertains to the way the algorithm uses the constraints, i.e. one at a time, with each reconstruction used as a prior for the next RE minimization. This is the first time the RE minimization with errors in the constraints has been used in image decompression even though it is common in spectrum estimation when there are errors in the correlation measurements.

Adaptive Solutions' CNAPS architecture is a parallel array of digital processors. This design features a Single-Instruction Multiple-Data (SIMD) stream architecture. The architecture is designed to execute on- chip learning for Artificial Neural Network (ANN) algorithms with unprecedented performance. ANNs have shown impressive results for solving difficult image processing tasks. However, current hardware prevents many ANN solutions from being effective products. The CNAPS architecture will provide the computational power to allow real time ANN applications. Because of the high parallelism of the architecture,it is also ideal for digital image processing tasks. This architecture will allow high performance applications that combine conventional image processing methods and ANNs on the same system. This paper gives a brief introduction to the CNAPS architecture, and gives the system performance on implementation of neural network algorithms, and conventional image processing algorithms such as convolution, and 2D Fourier transforms.

In this paper we introduce a new class of artificial neural network (ANN) models based on transformed domain feature extraction. Many optical and/or digital recognition systems based on transformed domain feature extraction are available in practice. Optical systems are inherently parallel in nature and are preferred for real time applications, whereas digital systems are more suitable for nonlinear operations. In our ANN models we combine advantages of both digital and optical systems. Many transformed domain feature extraction techniques have been developed during the last three decades. They include: the Fourier transform (FT), the Walsh Hadamard transform (WHT), the discrete cosine transform (DCT), etc. As an example, we have developed ANN models using the FT and WHT domain features. The models consist of two stages, the feature extraction stage and the recognition stage. We have used back-propagation and competitive learning algorithms in the recognition stage. We have used these ANN models for invariant object recognition. The models have been used successfully to recognize various types of aircraft, and also have been tested with test patterns. ANN models based on other transforms can be developed in a similar fashion.

Real—time visual tracking is a difficult problem requiring high speed processing. Neural networks have shown great potential for solving the problem. We have recently reported a two—layer Hopfield—Tank (HT) network implementation of a fast tracking algorithm capable of estimating displacements from a sequence of images [1]. In this paper we have taken a step further. That is, we have designed a custom analog VLSI chip capable of computing target motion parameters. To date, a number of researchers have constructed small—scale networks with less than 100 neurons as well as large—scale networks with up to 1000 neurons, both with feedback and feedforward architectures [2,3]. Small networks have been limited in size by VLSI real— estate considerations as well as its internal wiring. Large network have been constructed using "building block" paradigms. These building blocks are known as synapse and neuron. A synapse chip contains many synapse cells. Similarly, a neuron chip contains only neuron cells. A neuroprocessor is formed by connecting many of these chips together. Signals are routed off—chip. Hence, speed is slower than a single chip design. However, a large network with different architectures can be implemented with these building blocks. Our design is based on a single chip design. The single chip design was chosen in order to satisfy the high speed required for our application. This paper first briefly presents the overall architecture of the network. Individual cell design then follows together with its SPICE simulations.

In recent years neural networks have been used to solve some of the difficult real time character recognition problems. These SIMD implementations of the networks have achieve some success, but the real potential of neural networks are yet to be utilized. Several well known neural network architectures have been, modified, and implemented. These architecture are then applied to character recognition. The performance of these parallel character recognition systems are compared and contrasted. Feature localization and noise reduction are achieved using least squares optimized Gabor filtering. The filtered images are then presented to an FAUST based learning algorithm which produces the self- organizing sets of neural network generated features used for character recognition. Implementation of these algorithms on highly parallel computer with 1024 processors allows high speed character recognition to be achieved at a speed of 2.3 ms/image, with greater than 99% accuracy on machine print and 89% accuracy on unconstrained hand printed characters. These results are achieved using identical parallel processor programs demonstrating that the method is truly font independent. The back propagation is included to allow comparison with more conventional neural network character recognition methods. The network has one hidden layer with multiple concurrent feedback from the output layer to the hidden and from hidden layer to the input layer. This concurrent feedback and weight adjustment is only possible on a SIMD computer.

Anovel MIMD (Multiple Instruction Multiple Data) based architecture consisting of multiple processing elements (PE) has been developed. This architecture is adapted to real-time processing of sequences of different tasks for local image segments. Each PE contains an arithmetic processing unit (APU), adapted to parallel processing of low level operations, and a high level and control processor (HLCP) for medium and high level operations and control of the PE. This HLCP can be a standard signal processor or a RISC processor. Because of the local control of each PE by the HLCP and a SIMD structure of the APU, the overall system architecture is characterized as MIMD based with a local SIMD structure for low level processing. Due to an overlapped computation and communication the multiprocessor system achieves a linear speedup compared to a single processing element. Main parts of the PE have been realized as two ASICs in a 1.5 jim CMOS-Process. With a system clock rate of 25MHz, each PE provides a peak performance of 400 Mega operations per second (MOPS).