This paper describes a programmable processor for digital signal and image processing. The design has been implemented using the multilinear form approach. The multilinear form approach is used to map digital signal processing techniques onto array structures. The resulting structures are regular, modular, and highly parallel. This approach is systematic; therefore, it would be useful for logic synthesis. The application of this approach in digital signal processing (DSP) and numerical computations reduces the design time, which results in low design cost. In general, this approach produces various structures (semi-systolic, quasi-systolic, and pure systolic networks) that could be considered application specific array processors.

A method of pattern recognition using a three layered feedforward neural network is described. Experiments were carried out for handwritten katakana in a frame using neural network. Handwritten characters have varieties of scales, positions, and orientations. In a neural network, however, if the input patterns are shifted in position, rotated, and varied in scales, it does not function well. So we describe a method to solve the problems of these variations using three layered feedforward neural network. We used two kinds of moment values that are invariant for these variations. One is regular moments and the other is Zernike moment, which gives a set of orthogonal complex moments of an image known as Zernike moments. We also describe the problem of the structure of neural networks and the relation between the recognition rate and data sets for similar and different patterns.

There are many kinds of image processing that are essentially sequential. Raster scan is commonly used for such sequential operations, to scan images from left to right, and line by line. Another scanning, called the Peano scan, traverses an image from a pixel to its neighboring one and the direction frequently changes. This scan prevents from producing periodic patterns, which are sometimes observed in images transformed in raster scan line order. However, the Peano scan is applied to only square images, and the horizontal and vertical sizes must be a power of two. We present a new scanning, called a ternary scan, which has the same property as the Peano scan and can fit to any rectangular images. Application of the ternary scan to Floyd-Steinberg's halftoning is shown.

This paper presents a method of recognizing human face profile. The conventional methods of recognizing human face profile use the computer-derived fiducial marks, lines, angles, and other measures of profile outline as the components of characteristic vector. We use P-type Fourier descriptor as a characteristic vector of human face profile. It is shown that four P-type Fourier coefficients in the low frequency range can identify 65 face profiles, with the accuracy of 100%.

This paper proposes a new approach for designing robust pattern classifiers for human face images with the aid of a state-of-the-art 3-D imaging technique. The 3-D CG models of human faces are obtained using a new 3-D scanner. A database of synthesized face images simulating diverse imaging conditions is automatically constructed from the 3-D CG model of the subject's face by generating a series of images while varying the image synthesis parameters. The database is successfully applied to the extraction of a pair-wise discriminant that achieves higher class separability against real face images of two subjects acquired under disparate imaging conditions. The use of the 3-D CG model in training a classifier is shown to yield more accurate face recognition in the framework of 2-D image matching.

In this paper, we introduce a vision-based computer input device which has been built at the University of Georgia. The user of this system gives commands to the computer without touching any physical device. The system receives input through a CCD camera; it is PC- based and is built on top of the DOS operating system. The major components of the input device are: a monitor, an image capturing board, a CCD camera, and some software (developed by use). These are interfaced with a standard PC running under the DOS operating system.

A stable Gabor-type representation of an image requires that the Zak transform (ZT) of the reference function does not vanish over the fundamental cube. We prove that the discrete ZT of any symmetric set of reference data points has a zero. To overcome the computational problem, which is due to the zero plane generated by the ZT of the Gaussian reference function, the Gaussian is translated by a sub-pixel distance. We show that the absolute value of the minimum of the ZT of the Gaussian is a function of the sub-pixel distance of translation and that the optimum value of such translation is 1/2 pixel.

Recently, we have presented a recursive framework for noncausal Gauss Markov Random Fields (GMRF) defined on finite lattices. This framework readdresses the issue of recursiveness in 2-D signal processing, providing the means to attain the computational advantages of recursive processing without sacrificing the noncausality of the image model. We present here results on the application of this framework to two areas, image enhancement and image compression.

A set of neural lattices that are based on the central limit theorem is described. These lattices, generate in parallel, a set of multiple scale Gaussian smoothing of their input arrays. As the number of layers is increased, the generated kernels converge to ideal Gaussians with infinitely small error. In addition, the lattices can generate in parallel, a variety of multiple scale image operators such as: Canny's edge detectors, Laplacians of Gaussians, and Sine, Cosine, Fourier and Gabor transforms. It is also proved that any bounded signal, including sinusoidal kernels, can be approximated by a finite number of Gaussians with arbitrarily small error.

Grouping disconnected edgels and forming quantitative descriptions of structural entities are fundamental tasks in image understanding. Our newly developed grouping algorithm, the Distributed Hough Transform (DHT), is capable of performing both of these tasks. The DHT uses the principle of proximity weighted symmetry (PWS). PWS is based upon non- accidentalness, viewpoint invariance, and new probabilistic models of projected angles and distances which use the assumption that boundaries of man-made objects can be represented by circular arcs and straight lines. Experimental results show that the DHT is a robust algorithm.

Morphological algorithms for the parallel quantification and modeling of Gaussian image features are described. These algorithms are applicable to any image generation process which distributes the gray-scale values according to a normal distribution. Morphological operators can be applied to the image data to obtain two parameter images, one consisting of mean positions and amplitudes and the other consisting of estimates of standard deviations, which are then used to 'grow' (in parallel) the predicted Gaussian surfaces. Two methods to decompose and modulate the growth process (using the parameters images) are considered. One method grows the predicted Gaussian surface in terms of an approximating binomial distribution. The other method grows the desired Gaussian from smaller Gaussians of varying standard deviations.

Random noise interference in image pick-up and image transmission is an important restriction for vision systems. In this paper, interframe shift sampling (IFSS) transform has been used for diminishing noise interference and detecting weak image signal submerged by strong noise in communication systems.

The registration of shapes in two images appears as an indispensable step in many applications such as analysis of dynamic scenes involving shape matching. In this paper, we compute the corresponding shape specific points, which are invariant in rotation, scale, and translation, between two misaligned shapes by morphological operations. We find that erosion is enough to determine the specific points between two nontrivial binary shapes to be registered.

The principles of organization of biological visual systems provide a valuable inspiration to the design and construction of image acquisition systems. The adaptive sensitivity mechanism of human-like retina enables the acquisition of extremely wide range of light intensity (10¹³:1), much wider than that of conventional semiconductor photoreceptors (10³:1), while preserving image details and dynamically adapting to the image contents. In order to better design intelligent visual data acquisition and processing systems, a three- layer visual neural model would be proposed to exploit the attractive capability of human retina. Due to the limited range of photoreceptor, each photoreceptor in the first layer could improve its range by shifting the operating characteristics and adjusting the threshold along the intensity axis. The concept of light adaptation of photoreceptor is based on a combination of the photopigment bleaching and regeneration kinetics and the feedback neural inhibition mechanism. The horizontal cells which are located just below the photoreceptors form the second layer, which performs the spatially weighted averages of photoreceptor outputs and determines the settings of the light-adaptation parameters of neural inhibition mechanism. The third layer, which includes a number of bipolar cells, is used to perform the lateral inhibition that could do contrast and edge enhancement. This approach allows for perception of visible scene that is independent of changes in the overall level of illumination. A model of combination of light adaptation and lateral inhibition would be employed in designing the smart image acquisition systems and conducted to be verified in our HDTV Lab.

Mathematical morphology has been proven to be a very useful tool for image processing and computer vision. Gray scale and binary morphological operations are usually implemented by software. Software implementation is, however, inefficient and slow. To achieve real-time morphological processing, a new hardware implementation method is presented based on VLSI. The complexity of the implementation method is also discussed.

The block segmentation and block classification of digitized printed documents segmented into regions of texts, graphics, tables, and images are very important in automatic document analysis and understanding. Conventionally, the constrained run length algorithm (CRLA) has been proposed to segment digitized documents, however, it is space-consuming and time- consuming. The CRLA method must define some constrained parameters, so it cannot proceed automatically, and its performance may degrade significantly due to improper parameters. This paper proposes an efficient and effective method for document analysis, sequence connected segmentation and mapping matrix cell algorithm (SCSMMC). This method can analyze both simple and complex documents automatically and it need not define any constraint parameters. This method, which only needs one-reading image of document, can proceed completely and the techniques of segmentation, classification, labeling, and character segmentation proceed at the same time. The proposed document analysis method may also combine with the optical character recognizer to form an adaptive document understanding system.

A genetic algorithm approach to the design of dither matrices used in the dispersed-dot digital halftoning method is described. Digital halftoning techniques are used to render continuous- tone images on high resolution binary display devices. An important class of digital halftoning techniques are the ones that convert a gray level in an image into a binary pattern. The design of such patterns is important to reduce visible artifacts so as to render the image with higher fidelity. An approach to design based on shaping the spectrum of the dithering signal according to a model of the human visual system is presented. The effects on the search performance by the reproduction plan and genetic operators are demonstrated.

To improve the readability, image compression, and optical character recognition (OCR) system performance for two-tone (binary) text image data, we investigated morphological methods of image processing. We found them to be fast and effective not only with noise text images but with relatively noise-free images as well. Using morphology, we improved text image readability as judged in a blind test, increased compression ratio using CCITT Group 4, and reduced OCR error (cluster) rates in a commercial omnifront scanner.

Neural network architectures employing units called elementary velocity estimators to extract motion (direction and speed) from visual information have been considered. Design guidelines to determine the various parameters of units as different layers, given a set of specification, have been provided. In particular, the range of speeds that each unit or layer can accommodate has been explored. Furthermore, design methods have been introduced to determine the unit and the layer parameters for a specific range of speeds to be accommodated.

The paper presents the main features of a prototype image retrieval system, nicknamed Imagine. In this system, the image database is located in a site remote from the user workstation. The key issues in developing the prototype have been the response time and scalability, or the ability of maintaining a set of basic functionalities in a wide range of workstation performances and network digital rates. The paper focuses on the problems related to the image visualization process in a workstation with a limited number of reproducible colors. Three different approaches, split, shared, and generic colormap, are presented and compared.

A knowledge-based system is presented that has been designed to overcome the difficulties generally encountered in the extraction of structures from biomedical images. Different segmentation methods are opportunistically applied and parameter values are automatically controlled through the use of models, data, and progressive results. Detected structures are assigned fuzzy membership values related to the reliability of recognition results. The application of the system to microscopic images is described. Peculiar features of the system include a high degree of tolerance to parameter variations high flexibility and a reduced processing time.

The objective of this paper is to propose a syntactic formalism for space representation. Beside the well known advantages of hierarchical data structure, this underlying approach has the additional strength of self-adapting to a spatial structure at hand. The approach is called puzzletree because its recursive decomposition of an image results in a number of rectangular regions which in a certain order--like a puzzle--reconstruct the original image. The approach may also be applied to higher-dimensioned spaces. This paper concentrates on the principles of puzzletrees by explaining the underlying heuristic for their generation and outlining their use to facilitate higher-level operations like image segmentation or object recognition. Finally, results are shown and a comparison to conventional region quadtrees is done.

Morphological tau-openings are binary filters that are translation invariant, increasing, antiextensive, and idempotent. The Matheron representation for tau-opening filters show that they are always given by unions of elementary openings. Historically tau-openings have been employed to restore binary images corrupted by union noise. This paper investigates optimal restoration for the union noise model based upon the individual granulometric pattern spectra of the image and noise. It does so for a class of random-grain image and noise models.

The problem of the parallel implementation of a filter bank is discussed. An associative string processor architecture is described. A technique to generate filter banks with maximum localization is introduced. Simulations and complexity studies show that image sequence coding using a Gabor decomposition is possible at video rate.

Many texture-segmentation mechanisms take the form of an elaborate bank of filters. Commonly used within such mechanisms are the Gabor elementary functions (GEFs). While filter-bank-based mechanisms show promise and some analytical work has been done to demonstrate the efficacy of GEFs, the relationships between texture differences and the filter configurations required to discriminate these differences remain largely unknown. In this paper, we given analytical and experimental evidence that suggests that various types of discontinuities can occur at texture boundaries when appropriately 'tuned' GEF-based filters are applied to a textured image; thus, by applying a discontinuity-detection scheme (i.e., edge detector) to such a filtered image, one can segment the image into different textured regions.

This paper presents an analytical method to measure the performance through memory of a proposed parallel processing computer architecture for image processing applications. We developed an analytical model of our proposed architecture and a conventional architecture with respect to the system's local and main memory to measure and compare the performances of these two systems. From our model we can evaluate the performance of the proposed architecture in terms of processor utilization, number of busy memory modules and the fraction of the programs data structure resident in local memory. The main idea behind our proposed architecture is to carry out image processing work in a highly parallel manner so that the response time is shorter. The proposed architecture keeps both the processor and the memory system as busy as possible in order to obtain faster response time and proper utilization of the hardware compared to a conventional parallel processing architecture. Our proposed architecture consists of an array of processing elements (PEs), a system control unit (SCU), interconnection network and memory modules. Each PE contains two central processing units (CPU), one is responsible for the execution of all non-memory operations and the other is responsible for all memory operations. The overall response time of a job is faster because we divide the actual execution and the memory operation into two separate entities and carry them out concurrently. We also present an image processing algorithm suitable for the proposed architecture and analyze their performance compared to the conventional system.

The usefulness of the concept of a pseudo-inverse matrix to problems in signal processing is emphasized. It is shown how pseudo-inverse matrices can be used to describe the projection of a signal onto the subspace generated by a finite family of basis signals, which need not be linear independent. The connection to the theory of frames, the one-step POCS method, and Cunningham projection onto repetitive bases are explained, among others.

Directional selective unit, or velocity sensitive unit, is the mechanism in a vision system that responds to motion in a specific direction. This paper analyzes a model for directional selective unit proposed by Barlow and Leveck. This model employsthe inhibitory mechanism. Several parameters of the model are defined and examined. The analysis shows that the range of the response of the unit can be adjusted through the parameters, that the specific values of the parameters are not as important as the relative values of the parameters, and that a family of velocity-sensitive units with different ranges can be used to achieve velocity-specific detection.

The paper presents a unified texture model, which is applicable to a wide variety of texture types found in natural images. This model leads to the derivation of texture analysis and synthesis algorithms designed to estimate the texture parameters and reconstruct the original texture field from these parameters. The model is highly motivated by findings about human vision. The texture field is assumed by a mixed spectral distribution. On the basis of a 2-D Wold like decomposition for homogeneous random fields, the texture field is decomposed into a sum of two mutually orthogonal components: a purely-indeterministic component and a deterministic component. The deterministic component is further orthogonally decomposed into a harmonic component, and a generalized-evanescent component. The purely- indeterministic component is represented by a 2-D, non-symmetrical-half-plane, finite support AR model. The harmonic random field is a sum of 2-D harmonic components of random amplitude and phase. The generalized evanescent field consists of a countable number of wave systems all traveling in directions of rational tangent, and all modulated by 1-D purely- indeterministic processes in the orthogonal dimension. Both analytical and experimental results show that the deterministic components should be parametrized separately from the purely- indeterministic component. The model is very efficient in terms of the number of parameters required to faithfully represent textures. Reconstructed textures are practically indistinguishable from the originals.

This paper proposes a background region removal method that extracts facial video images for their later superimposition onto retrieved images held in a video filing system. The method is robust against light intensity fluctuation. The background is eliminated from the camera signal by a combination of differential signal thresholding, filtering, and masking. Current thresholding methods are somewhat unstable because they fail to handle fluctuations in light intensity. The proposed method is very stable because it uses two assymetric thresholds in relation to the origin. The method is quite simple and fast. It consists of four steps: storing the background image in the background memory; subtracting the background image from the camera output; selecting two appropriate thresholds; and thresholding to generate a key signal as in the chroma-key technique. The key part of the method is threshold selection. Thresholds are selected by using a Gaussian and constant function approximations of the differential signal's histogram so as to eliminate histogram deterioration caused by light intensity fluctuations.

This paper describes how a family of images of varying levels of contrast can be efficiently calculated and displayed. We show that a contrast space defined in terms of histogram equalization can be specified in terms of two parameters: region size and histogram blurring level. Based on these observations, we describe one exact algorithm and two efficient algorithms for computing sequences of images within this contrast space. These precomputed images can be displayed and interactively explored to examine image features of interest. By working with a family of contrast enhanced images the difficult task of selecting the single level of contrast enhancement appropriate for a particular image is avoided, increasing the usefulness of low contrast images.

Decomposing morphological structure element into Minkowski sum of several small ones is very useful for fast implementation of morphological operations and important for multiscale systems. This paper presents some theoretical results on decomposition of digital structure element, such as geometrical constraints, singularity, compatibility, and decomposability, etc., which is very different from that in continuous space. Based on those, methodology for decomposition will be proposed, including approximate decomposition, correction of singularity, and etc. These results will be used in decomposition into four and eight neighborhood configurations and series decomposition of multiscale structure sequence in the paper.

A consistent, reliable method of quantifying diabetic retinopathy is required, both for patient assessment and eventually for use in screening tests for diabetes. To this end, an algorithm for determining the degree of venous beading in digitized ocular fundus images has been developed. A parallel implementation of the algorithm has also been investigated. The algorithm thresholds the fundus image to extract vein silhouettes. Morphological closing is used to fill any anomolous holes. Thinning is used to determine vein centerlines. Vein diameters are measured normal to the centerlines. A frequency analysis of vein diameter with distance along the centerline is then performed to permit estimation of veinous beading. For the parallel implementation, the binary vein silhouette and the vein centerline are rotated so that vein diameter may be estimated in one direction only. The time complexity of the parallel algorithm is O(N). Algorithm performance is demonstrated with real fundus images. A simulation of the parallel algorithm is used with actual fundus images.

A multi-task neuro-vision parameter which performs a variety of information processing operations associated with the early stages of biological vision is presented. The network architecture of this neuro-vision processor, called the positive-negative (PN) neural processor, is loosely based on the neural activity fields exhibited by thalamic and cortical nervous tissue layers. The computational operation performed by the processor arises from the strength of the recurrent feedback among the numerous positive and negative neural computing units. By adjusting the feedback connections it is possible to generate diverse dynamic behavior that may be used for short-term visual memory (STVM), spatio-temporal filtering (STF), and pulse frequency modulation (PFM). The information attributes that are to be processes may be regulated by modifying the feedforward connections from the signal space to the neural processor.

A new method for video signal processing is described in this paper. The purpose is real-time image transformations at low cost, low power, and small size hardware. This is impossible without special hardware. Here generalized digital differential analyzer (DDA) and control memory (CM) play a very important role. Then indentation, which is called jaggy, is caused on the boundary of a background and a foreground accompanied with the processing. Jaggy does not occur inside the transformed image because of adopting linear interpretation. But it does occur inherently on the boundary of the background and the transformed images. It causes deterioration of image quality, and must be avoided. There are two well-know ways to improve image quality, blurring and supersampling. The former does not have much effect, and the latter has the much higher cost of computing. As a means of settling such a trouble, a method is proposed, which searches for positions that may arise jaggy and smooths such points. Computer simulations based on the real data from VTR, one scene of a movie, are presented to demonstrate our proposed scheme using DDA and CMs and to confirm the effectiveness on various transformations.

In studying high efficiency color image coding and image processing, it is important to segment several regions that represent real objects. By performing this region segmentation, we can establish the structural description using the characteristic information of regions. To segment the image into several characteristic regions, we adopt the clustering algorithm in the Uniform Lightness Chromaticness Scale System and the merging process based on the measure of Godlove's color difference.

A new method for detecting dominant points is presented. It does not require any input parameter, and the dominant points obtained by this method remain relatively the same even when the object curve is scaled or rotated. In this method, for each boundary point, a support region is assigned to the point based on its local properties. Each point is then smoothed by a Gaussian filter with a width proportional to its determined support region. A significance measure for each point is then compared. Dominant points are finally obtained through nonmaximum suppression. Unlike other dominant point detection algorithms which are sensitive to scaling and rotation of the object curve, the new method will overcome this difficulty. Furthermore, it is robust in the presence of noise. The proposed new method is compared to a well-known dominant point detection algorithm in terms of the computational complexity and the approximation errors.

We present two robust adaptive median filters with variable window size which are capable of removing a mixture of positive and negative impulse noise while preserving sharpness of an edge. In the first case, we assume that each pixel at (i J ) is corrupted by an impulse with probabilitype independent of other pixels corrupted or not. The impulse corrupted pixel takes on the minimum pixel value smm with probability q or the maximum pixel value mwith probability 1 when the original pixel s is corrupted by a negative or a positive impulse, respectively. Let {x } be the noise corrupted image. Then I e,1 with Pe xii = 1 with 1 The RAMF algorithm is based on a test for the presence of an impulse at the center pixel followed by a test for the detection of residual impulse in the median filter output. In the second model, the noise corrupted pixel is x =s +n1 , where n1 is iid impulsive noise having Laplacian, or Cauchy, or a mixture of Gaussian and Cauchy distributions. The SAMF algorithm in this instance detects the width of the impulse and adjusts the window accordingly until the noise is eliminated. These algorithms were tested on standard images. The RAMF is shown to be supenor to the nonlinear mean L filter[1] while the SAMF is better performing and simpler than [in's adaptive scheme[3].

The extraction of component objects from images is fundamentally important for image synthesis. In TV program production, one useful method is the Video-Matte technique for specifying the necessary boundary of an object. This, however, involves some manually intricate and tedious processes. A new method proposed in this paper can reduce the needed level of operator skill and simplify object extraction. The object is automatically extracted by just a simple drawing of a thick boundary line. The basic principle involves a thinning of the thick boundary line binary image using the edge intensity of the original image. This method has many practical advantages, including the simplicity of specifying an object, the high accuracy of thinned-out boundary line, its ease of application to moving images, and the lack of any need for adjustment.

A weighted outer product learning (WOPL) scheme for associative memory neural network is presented in which learning orders are incorporated to the Hopfield model. WOPL can be guaranteed to achieve correct recall of some stored datums no matter whether or not they are stable in the Hopfield model, and whether the number of stored datums is small or large. A technically sufficient condition is also discussed for how to suitably choose learning orders to fully utilize WOPL for correct recall of as many stored datums as possible.

High quality color picture painting system may be one of the most interesting problems for various application fields such as infant education, natural color picture animation synthesis, and hobby works. Picture signal can be divided into luminance data and chrominance data. Luminance data is scalar and chrominance data is vector. And then, in those signals, the parts of discontinuity is considered as outlines of the picture. In order to process high quality color pictures it is necessary to consider the properties of outline and region as features of the picture. Thus, we need the system design that uses point, line, and region features for frame structure and manipulates luminance and chrominance data as attribute values. Also, we must consider duality of picture processing and painting. Moreover, for software productivity, reliability, and extensibility, it is wrong to implement directly by using general language such as C. In the process of software production, we want to emphasize the needs of the Application Specific Language. The Window-based Elaboration Language for Picture Processing and Painting (WELL-PPP) is developed as an example of the Application Specific Language. In the developing process, WELL-PPP plays an important role for actual implementation of software process that includes defining a set of data, operations and windows, and many objects. WELL-PPP also defines the executing process of operations. Exchange of data between windows is managed by standard protocol of WELL-PPP. WELL- PPP has a close relation with the TELL system. This paper describes the processes of software production and color picture processing and painting. This software process is suitable for collaborated works.

In this paper we propose two algorithms for the restoration of images based on the bispectrum. The bispectrum of a signal is the Fourier transform of its triple correlation. While second- order statistics (e.g., correlation function, power spectrum, etc.) do not provide any information about the phase of the signal, third-order statistics (e.g., triple correlation, bispectrum, etc.) allow the recovery of the phase of the signal. We propose two algorithms for estimating the magnitude and the phase of the image, where the ambiguity due to the use of the principal value of the phase component is taken into account. Image lines are used in our experiments to test the effectiveness of the proposed algorithms.

Two adaptive approaches for nonstationary filtering of image sequences are presented and experimentally compared. According to the first approach, a recursive spatio-temporal motion- compensated (MC) estimator is applied to the noisy sequence that adapts to the local spatial and temporal signal activity. A separable 3-D estimator is proposed that consists of three coupled 1-D estimators; its input is the noisy image plus additional signals that contain spatial information provided by a simple edge-detector or temporal information provided by the MC backward difference (registration error). The steady-state gain and the parameters of this separable estimator are computed by closed form formulae, thus allowing a very efficient implementation. According to the second approach, the noisy signal is first decomposed into a stationary and a nonstationary part based on an estimate of its local mean and deviation. A minimum variance estimator of the local mean and deviation of the observed signal is used. After the current mean is subtracted from the observed signal and the signal is normalized by using the current deviation, a relatively simple noise filter is used for filtering the stationary part. The above methods are applied to the filtering of noisy video-conference image sequences for various levels of noise. Both methods show a very satisfactory performance taking into consideration their simplicity and computational efficiency.

This paper describes a new address generation unit (AGU) architecture for video signal processing. The proposed AGU has several sophisticated addressing modes obtained by analyzing many video signal processing algorithms, and can generate image memory addresses needed in video signal processing. This AGU's architecture was employed in a DSP being developed at present. In this paper, first of all, the architecture of the DSP and the role of the AGU are presented. Furthermore, the details of this AGU, such as the control method, addressing modes, and operations of the AGU in some typical applications are described.

This paper presents a method for natural texture analysis in a multiscale context. The texture signature is taken to be an anisotropic fractal feature related to the fractal dimension. By computing this signature in different orientations and through several pyramid levels, an attribute vector is obtained. The experimental results of texture segmentation by using Bayesian classifier show the method is efficient.

The resolution of computerized tomographic (CT) images is limited due to the bandlimited reconstruction process and the use of smoothing windows that eliminate high spatial frequencies from the image. A new method is presented that restores these missing spectral components and thus increases the resulting spatial resolution of the image. The new method based on the row action projection (RAP) algorithm is computationally efficient and facilitates local adaptation of the projection operators. The local mean value as well as minimum and maximum bounds are used as constraints. The method is proposed to provide a zoom-in capability which yields a high resolution estimate of a specified region of the image. The zoom-in feature could be of great utility in medical and commercial applications of tomographic image reconstruction. Computer simulations demonstrate the new method to be very effective in recovering high order spectral components of designated regions of the reconstructed image.

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

In the motion picture coding, postprocessing plays an essential role to reduce the blurring the noise of the reconstructed images. In this paper, we present a motion dependent median filtering which combines the spatial and temporal domain operations and motion estimation to preserve sharpness of video sequence. The experimental results show that our proposed filters can achieve better objective results and higher subjective qualities than the other three- dimensional filters.

We propose a fully automatic system that reconstructs a 3-D tree structure based on two images taken from mutually orthogonal directions. The system comprises two major components, the high-level knowledge base and the low-level signal processing algorithms. Computational algorithms are responsible for tracking segments in each image, representing 2- D segments with directed graphs, and reconstructing 3-D segments from matching 2-D segment pairs. Knowledge of epipolar constraint, segments geometry, and connectivity are used to resolve the problem of matching segments between two views. Interaction between segmentation and matching is facilitated by uncertain reasoning. The knowledge base is represented by production rules which are implemented in the CLIPS shell. The computational algorithms are coded in the C language. The performance of the system is evaluated with images of a tree-like model. The total processing time for reconstruction of this tree-like model with 256³ resolution is 14 seconds on a workstation. The results show that the system is capable of correctly delineating an object's 3-D tree structure. This study demonstrates a non model-based approach to two-view reconstruction of tree-like structures.

A number of training algorithms for neural networks are based on the 'competition' learning method. This is regarded as an adaptive process for tuning neural networks to specific features of input. The responses from the neural network, then, tend to become localized. However, a shortcoming of this model is that some neural units can remain inactive. Since a neural unit never learns unless it wins, it is possible that some of the neural units are always outperformed by others, and therefore never learn. This paper presents a new unsupervised learning algorithm, less-interclass-disturbance learning (LID), which deals with the limitations of the simple competitive neural network. The main idea of the method is that it reinforces the competing neurons in such a way as to prevent the weights from 'fooling around.' A new compound similarity metric is introduced in this algorithm to reduce the interclass disturbance during the training process. The behavior of this algorithm was investigated through computer simulations. It is shown that LID learning is quite effective.

In this paper, we develop a bidirectional associative memory (BAM) based neural network to achieve high-speed partial shape recognition. To recognize objects which are partially occluded, we represent each object by a set of landmarks. The landmarks of an object are points of interest relative to the object that have important shape attributes. To achieve recognition, feature values (landmark values) of each model object are trained and stored in the network. Each memory cell is trained to store landmark values of a model object for all possible positions. Given a scene which may consist of several objects, landmarks in the scene are first extracted, and their corresponding landmark values are computed. Scene landmarks values are entered to each trained memory cell. The memory cell is shown to be able to recall the position of the model object in the scene. A heuristic measure is then computed to validate the recognition.

In this paper, a new pattern recognition scheme is proposed by the authors, which features compressing a huge input vector into a tiny one and catching the characteristics of the input vector efficiently. The development of this scheme is based on a theory of class 2 dynamical systems, where the class 2 dynamical system is defined by the authors. An approach using (omega) -Orbit Finite Automata developed by the authors is a special class of this method. This scheme has two stages, encoding and quantization. The encoding procedure stores an input vector in an attractor of a class 2 dynamical system. The quantization procedure divides the parameter space of the class 2 dynamical systems inferred at encoding stage. A retrieval algorithm for (omega) -OFA and several inference algorithms of class 2 dynamical system from a given input vector are introduced.

A new algorithm for multiscale morphological descriptions of binary digital regions is given. From a bound set of pixels in digital space it extracts a growing sequence of subsets approximating it; each approximation is obtained from the previous one by addition of the opening by one structuring element chosen in a finite family. Corresponding to this sequence of subsets is a compact coding specifying for each step the structuring element used, and a set of positions for this structuring element, which is sufficient to generate the difference between the current approximation and the previous one.

This paper presents a pattern recognition system that self-organizes to recognize objects by shape. The images are processed using a log-polar transformation that maps rotations and magnifications into representative translations. The systems then uses a multistaged hierarchical neural network that exhibits insensitivity to translations in representation space, which corresponds to rotations and scalings in the image space. The network's three layers perform the functionally disjoint tasks of preprocessing (dynamic thresholding), invariance (position normalization), and recognition (identification of the shape). The Preprocessing stage uses a single layer of elements to dynamically threshold the grey level input image into a binary image. The Invariance stage is a multilayered neural network implementation of a modified Walsh-Hadamard transform that generates a representation of the object that is invariant with respect to the object's position, which maps back to an invariance to rotational orientation and/or size. The Recognition stage is a modified version of Fukushima's Neocognitron that identifies the normalized representation by shape. The resulting network can successfully recognize objects that have been rotated, scaled, or a combination of both. The network uses a small number of fairly simple elements, a subset of which self-organize to produce the recognition performance.

Development of 'intelligent' safety systems using computer vision is discussed. Robots usually carry out the same movements at each work cycle. In this situation we can modelize at each sequence the robots' position. The spatio-temporal model of the robotic scene's behavior obtained by learning accelerates the comprehension of the observed image. The synchronized model is compared with the real scene at each sequence. The prediction of future movements of moving objects obtained with our model facilitates the 'region of interest' generation in the studied scene and, in the same way, helps decisions of the safety device. The adaptation of the surveillance task to the changes of illumination reduces false alarms.

This paper describes a method of reproducing three-dimensional face models using multi-view images for a virtual space teleconferencing system that achieves a realistic visual presence for teleconferencing. The goal of this research, as an integral component of a virtual space teleconferencing system, is to generate a three-dimensional face model from facial images, synthesize images of the model virtually viewed from different angles, and with natural shadow to suit the lighting conditions of the virtual space. The proposed method is as follows: first, front and side view images of the human face are taken by TV cameras. The 3D data of facial feature points are obtained from front- and side-views by an image processing technique based on the color, shape, and correlation of face components. Using these 3D data, the prepared base face models, representing typical Japanese male and female faces, are modified to approximate the input facial image. The personal face model, representing the individual character, is then reproduced. Next, an oblique view image is taken by TV camera. The feature points of the oblique view image are extracted using the same image processing technique. A more precise personal model is reproduced by fitting the boundary of the personal face model to the boundary of the oblique view image. The modified boundary of the personal face model is determined by using face direction, namely rotation angle, which is detected based on the extracted feature points. After the 3D model is established, the new images are synthesized by mapping facial texture onto the model.

An algorithm for quality inspection of characters printed on chip resistors is presented in this paper. Chip resistors are extremely small electronic parts used in small electronic devices, whose production have increased because of their ease of installation. To decrease production costs and ensure uniformity of quality, chip resistors are produced in automated manufacturing plants. Appearance inspection, however, still depends on visual inspection. It is quite difficult, however, to examine the printing quality of the characters printed on the chip resistor, because the characters are very small, reaching the limit of printing technology. Another difficulty is that, at present, there is no standard of quality. Only a few manuals exist describing defects. These have been written for use in individual factories and differ from factory to factory. Furthermore, even in one factory the criteria vary from one person to another. An automatic quality inspection system for the characters printed on such small chip resistors is urgently required.

While temporal properties of the visual system have been the subject of extensive research in psychology, many computational theories are based on steady-state behavior. For example, Marr's theory requires early measurements to be instantaneous. Furthermore, optimizational type approaches to perception are designed around properties of equilibria, and very little attention is devoted to the relevance of trajectories to perceptual experience. Electrophysiological findings however show that visual neurons such as retinal ganglion cells possess strong transient components. Therefore, a fundamental issue in perceptual sciences is the understanding of the relevance of these transient components to visual perception. This study claims that adaptive, nonmonotonic transient properties of early visual units are crucial components in visual processing. An extra-retinal feedback on-center off-surround anatomy is proposed to sharpen the 'blurred output' from the retinal level. Based on theoretical studies of pattern transformation properties of recurrent networks for sustained inputs we propose a global model (including retina and extra-retinal areas) of visual processing where a reset from transient ganglion cells of the retina prevent smearing for moving images. The model provides a theoretical link between hyperacuity (achieved by denser extra-retinal packing and nonlinear contrast enhancement) and visual masking (resulting from inter-layer and intra-channel inhibition mechanisms).

A novel method for reconstruction of offset field quincunx coded color image sequences by processing signal components as vectors is presented. A multidimensional weighted vector median interpolation filter is introduced. The suggested interpolation method provides perfect reconstruction of still image areas and has good properties in moving areas of the sequence without any motion information. The performance of the proposed methods is tested using a video sequencer.

This paper presents methods for video browsing and extracting information from video by measuring brightness data. Cut breaks are detected by measuring areas in which inter-frame difference occurs. We propose a video browsing tool (VBTool) that creates a compressed video containing all important scenes. We utilize two factors related to pixel brightness to allow the automatic generation of the compressed video.

An audio-graphic teleconferencing system has been developed that uses ordinary personal computers (PCs) interconnected over a basic rate (2B+D) ISDN line. The system supports high-speed transmission of 200-dpi resolution documents read by an optical scanner and presented on the displays of the conference participants. While looking at the same material, the conferees can interactively converse and make handwritten notations for all the participants to see on the document via a LCD tablet. This paper describes the configuration and performance of the system, focusing mainly on the ISDN-based multi-media transmission method and the method of reducing and enlarging binary images.

In this paper we propose a computationally efficient multiframe Wiener filtering algorithm, called the cross-correlated multiframe (CCMF) Wiener filtering, for restoring image sequences that are degraded by both blur and noise. The CCMF approach accounts for both intraframe (spatial) and interframe (temporal) correlations by directly utilizing power and cross-power spectra of the frames. We propose an efficient implementation of the CCMF filter which requires the inversion of only N X N matrices, where N is the number of frames used in the restoration. Furthermore, is it shown that if the auto and cross-power spectra are estimated based on a three-dimensional (3-D) multiframe autoregressive (AR) model, no matrix inversion is required. We present restoration results using the proposed approach, and compare them with those obtained by restoring each frame independently using a single-frame Wiener filter. In addition, we provide the results of an extensive study on the performance and robustness of the proposed algorithm in the case of varying blur, noise, and power and cross- power spectra estimation methods using different image sequences.

In this paper, we propose a method for low-level unsupervised image segmentation via texture recognition and feature space clustering. The texture measure is based on the computation of n-tuple features of gray level values within the co-occurrence operator. These features are extracted from small local areas of the image. The strategy results in a feature vector transformation of the image. Self-evolving clustering is then used to group these feature vectors into clusters of homogeneous textured regions. The method as presented is applied to, and shown to be capable of, segmenting natural texture image composites. The method is computationally simple and can be implemented in hardware for real-time operation.

The luminance emitted from a cathode ray tube, (CRT) display is a nonlinear function (the gamma function) of the input video signal voltage. In most analog video systems, compensation for this nonlinear transfer function is implemented in the camera amplifiers. When CRT displays are used to present psychophysical stimuli in vision research, the specific display nonlinearity usually is measured and accounted for to ensure that the luminance of each pixel in the synthetic image properly represents the intended value. However, when using digital image processing, the linear analog-to-digital converters store a digital image that is nonlinearly related to the displayed or recorded image. This paper describes the effect of this nonlinear transformation on a variety of image-processing applications used in visual communications.

This paper describes an adaptive (self-learning) approach to object identification developed during the last two years. This approach combines an adaptive neural network with a model- based approach to object identification. It is based on the Maximum Likelihood Adaptive Neural System (MLANS), which has a capability for self-learning invariant features and symbolic patterns.

Two methods for detecting dim, unresolved target tracks in infrared imagery are presented. Detecting such targets in a sequence of noisy images is very challenging from the standpoint of algorithm design as well as detection performance evaluation. Since the signal-to-noise ratio per pixel is very low (a dim target) and the target is unresolved (of spatial extent less than a pixel), one must rely on integration over target tracks which span over many image frames. In addition, since there is a large amount of uncertainty as to the pattern and location of target tracks, good algorithms must consider a large number of possibilities. The first method is based on a generalization of the Hough transform-based algorithm using the Radon transform. The second approach is an extension of a detection theory algorithm to 3-D. Both algorithms use a 3-D volume of spatial-temporal data.

The present paper investigates motion compensated temporal filtering for noise removal in medical image sequences. A first order recursive filter that adapts automatically to the quality of motion estimation is tested, and two extensions of the algorithm are suggested. One is a combination with a efficient, edge preserving spatial filter, and the other an extension of second order recursive filtering with fa one frame delayed output. The latter is achieved without demands for extra motion estimates. The extensions give improvements in output signal-to-noise ratio, both if used separately, and more so if used in combination.

The computation of an estimate of the autocorrelation function from available data enters a great number of signal processing applications and typically represents the bulk of the computation time required in each application. This work investigates frequency domain techniques for the evaluation of the autocorrelation of multidimensional signals: in particular, the extension of Rader's algorithm for 1D signals is considered. The bidimensional case is treated in detail because it is of special interest for applications and because the reasoning used can be readily applied to higher dimension signals. The direct extension of Rader's algorithm to the multidimensional case is not optimal with respect to the choice of the subblock dimension, unlike in the one dimensional case: a modified algorithm is proposed that allows further computational savings and is particularly attractive for the data organization. The computation time required by frequency domain techniques is evaluated in detail. The analysis confirms that the proposed frequency domain techniques lead to significant computation time savings.

The measuring and display system for a marathon TV program employing a real-time image processor and a fast graphic processor has been developed. The system consists of three parts: the step frequency subsystem; the automatic TV camera tracking subsystem; and the computer graphic display subsystem. Actually it is the first trial which could track the motion of a runner's face and obtain the step frequency and display the runner's motion at a real-time rate in the sports program. In the step frequency subsystem, the step frequency for a marathon runner is detected by applying two different methods: measuring the area corresponding the figure of a runner and tracking the motion vector of a runner's face. By applying the result of detecting the motion vector, the automatic TV camera tracking subsystem is implemented. In the computer graphic display subsystem, the data for the position of each joint in the runner's leg are prepared beforehand. By applying the step frequency to those prepared data, the runner's motion is drawn out as an 3-D animation.

Images with motion parallax generated by CG should be displayed, based on the user's viewpoint, without delay. In this paper we examine the influence of delay time on images with motion parallax, both subjectively and via the task, by means of a device for 3-D manipulation in virtual space, and we measure the delay time in which the user can easily perceive an image with motion parallax. We also present a method to generate CG image at high speed in the allowable time by selecting from a database of different detailed models one model based on the distance between the viewpoint and object.

A framework for an optimal analysis of a large class of patterns deformed by affine transformation groups is presented. This approach is based on the properties of the Fourier cross-correlation and Lie groups theory. Group properties such as homogeneity, symmetry, and isometry are utilized naturally. In particular, we consider the important groups of similarities and rigid motion in plane and space. The method is general to any object functions: picture, shape, curve, etc.

A new approach to the robust recognition of objects is presented. The fundamental picture primitives employed are local orientations, rather than the more traditionally used edge positions. A simple technique of feature-matching is used, based on the accumulation of evidence in binary channels (similar to the Hough transform) followed by a weighted non- linear sum of the evidence accumulators (matched filters, similar to those used in neural networks). By layering this simple feature-matcher, a hierarchical scheme is produced whose base is a binary representation of local orientations. The individual layers represent increasing levels of abstraction in the search for an object, so that the object can be arbitrarily complex. The universal algorithm presented can be implemented in less than 100 lines of a high-level programming language (e.g., Pascal). As evidenced by practical examples of various complexities, objects can be reliably and robustly identified in a wide variety of surroundings.

A computer-assisted video technique is presented for rapidly and accurately gathering, storing, and depicting autoradiographic images, three-dimensional structure think specimens, autoradiographic images, etc. It is possible to study biochemical processes at multiple levels of the neuraxis. The technique employs a real-time digitizing frame store, color photographic putoff shadow subsystem, a microscope-graphic tablet input subsystem, and a microcomputer. Several significant advances described in this report include B-spline processing, algorithm of clockwise/counter-clockwise double coordinates, chain code, and linearization of gray values.

A new method for reconstructing 3-D shape from CT images is proposed in this paper. This approach recovers the original appearance of the organ by applying an evolutionary process to generate the interpolated contours for filling gaps between adjacent slices. It utilizes global shape information and prevents complicated matching with local features. It is powerful in handling CT images with large or unequal distance between slices. The morphological operations with suitable 3-D structuring elements, which can be implemented with parallel hardware, are used to suppress small fluctuations on the shape. Our new method provides a better visualization tool in clinical CT applications.

We present a method for measuring the three-dimensional orientation of planar surfaces. We derive a model relating the spatially varying instantaneous frequency of the image texture to the instantaneous frequency of the surface texture, to the orientation of the surface, and to the parameters of the imaging system. We measure the localized frequency at each image point with Gabor wavelets and use it to solve for the surface orientation according to the model. The method does not require the extraction of discrete texture elements. The algorithm has a mean error of about 5 degrees in the measured slant and tilt on a test set of 12 real-world surfaces.

This paper presents an overview of methods for recognition of omnifont printed Roman alphabet characters with various fonts, sizes and formats (plain, bold, etc.) from OCR system perspectives. First, it summarizes the current needs for optical printed character recognition (OPCR) in general, and then describes its importance for conversion between paper and electronic media. Current status of commercially available software and products for OPCR are briefly reviewed. Analysis indicates that the challenge we face in OPCR is far from being solved, and there is still a great gap between human needs and machine reading capabilities. Second, OPCR systems and algorithms are briefly reviewed and compared from the context of digital document processing for the following four stages: preprocessing of images, segmentation, recognition, and post-processing. Finally, possible research directions to improve the performance of OPCR systems are suggested, such as using an approach based on the combination of template matching and varieties of feature-based algorithms to recognize isolated characters, the use of multilayered architectures for OPCR, and parallel processing- based high-performance architectures.

In this paper, two control point-based image compression methods are presented. In the first method, the encoding is based on the roughness of the surface defined by the gray levels of the image. The second method utilizes a new distance transform, called the Distance Transform on Curved Space (DTOCS). The related compression ratios of both of the methods are very good. Also the computations needed are quite simple and require only a short processing time. The study includes the investigation of properties of the control point-based interpolation.

Most classification techniques either adopt an approach based directly on the statistical characteristics of the pattern classes involved, or they transform the patterns in a feature space and try to separate the point clusters in this space. An alternative approach based on memory networks has been presented, its novelty being that it can be implemented in parallel and it utilizes direct features of the patterns rather than statistical characteristics. This study presents a new approach for pattern classification using pseudo 2-D binary cellular automata (CA). This approach resembles the memory network classifier in the sense that it is based on an adaptive knowledge based formed during a training phase, and also in the fact that both methods utilize pattern features that are directly available. The main advantage of this approach is that the sensitivity of the pattern classifier can be controlled. The proposed pattern classifier has been designed using 1.5 micrometers design rules for an N-well CMOS process. Layout has been achieved using SOLO 1400. Binary pseudo 2-D hybrid additive CA (HACA) is described in the second section of this paper. The third section describes the operation of the pattern classifier and the fourth section presents some possible applications. The VLSI implementation of the pattern classifier is presented in the fifth section and, finally, the sixth section draws conclusions from the results obtained.

In this research, the use of an approximate method to derive edge localization parameter (edge width) at the output of the IDS filter is shown. This parameter is used to develop an algorithm for image reconstruction of the response of the IDS filter. Simulation of this algorithm on real images is illustrated.

Shape analysis is important in classification where deviation of a contour from a model are measured. In pathology, deviation of nuclear contour from an ellipse is sometime useful in cell clas— sification. This paper presents a method based on Fourier series expansion, to determine the best fit ellipse to a given contour by minimizing the difference in the ellipticity factors. We show that the geometric mean should be used for the contour center instead of the conventional centroid. The method is applicable to multi value contours. Examples are presented.

Presented in this paper is a new approach that can generate the shaded space-filling molecular models more realistically with the light source being in any front direction instead of behind the viewer that is commonly used. Decrease of the intensity with distance is also taken into account to improve realism. Franklin's linearization principle is modified and employed to apply to the situation of intersected spheres and the algorithm is thus efficient.

The main objective of this paper is to discuss a relatively new approach to the synthesis and analysis of shape. Specifically, we propose to characterize shape by means of random sets. The proposed approach may allow us to develop new shape synthesis and analysis procedures by combining probability theory with mathematical morphology. This combination has the potential of allowing modeling and analysis of shapes of both geometric and random structure.

A method for registering terrain range images is described. A smooth surface is fit to range data to provide a continuous mapping to second order surface intrinsics. Level sets of such intrinsics are found as features. These features enable two accurate means of comparison: arclength and a circular correlation of the distance from the discrete points on the level set to its centroid. These two methods of feature matching allow a matching algorithm with computational complexity O(N log N). The solution of the transformation from the matches employs a weighted least squares technique and removes matching errors by cross validation. The method is tested on simulated range images and results are given.

Prosthetic socket design is the most important aspect of the fit of a lower extremity prosthesis. The comfort and mobility of wearing a prosthetic socket mainly depends on the design of weight bearing characteristics of the socket. The weight bearing characteristics of a socket are determined by the relative positions among the tissues (such as bones, muscles, and fat) of the residual limb and the wall of the socket. Therefore, socket fitting cannot be guaranteed if the internal structural information of a residual limb is not made available to the prosthetist. Current prosthetic socket design and manufacture processes are disadvantaged by the inherent difficulties in determining the weight bearing characteristics of a socket due to the lack of crucial information about the internal structure of the residual limb, such as bone position, and muscle, and fat distribution. This disadvantage can be overcome through the use of an ultrasound imaging system and a computer-aided socket design system. The discussion of the complete system is out of the scope of this paper. An ultrasound imaging algorithm which provides the external shape of the limb, the bone position, and the texture pattern of the soft tissues within the limb is presented in this paper. The algorithm using the Compound B-scan imaging principle combined with geometrical transformation and statistical information faithfully measures and reconstructs with topographical shape and internal structure information of a residual limb.

To characterize how the human visual system responds to spatial patterns, a 'black box' method was adopted, in which visual evoked potentials (VEPs) were taken as outputs and visual patterned stimuli were taken as inputs. The stimuli were gratings whose function profiles were weighted Hermite polynomials (WHPs). A model, mathematically analogous to harmonic oscillators in quantum mechanics, was developed to describe the black box and the quantitative relationships between the VEPs and the WHPs.

This paper introduces the class of thresholded min-max functions and studies their learning under the probably approximately correct (PAC) model introduced by Valiant. These functions can be used as pattern classifiers of both real-valued and binary-valued feature vectors. They are a lattice-theoretic generalization of Boolean functions and are also related to three-layer perceptrons and morphological signal operators. Several subclasses of the thresholded min- max functions are shown to be learnable under the PAC model.

In this paper generalize weighted order statistic (GWOS) filters are introduced. As a subclass of generalized stack filters, GWOS filters can be implemented using a sorting operation in the real domain. Based on the relationship between GWOS filters and neural networks, two efficient adaptive algorithms are derived for finding optimal GWOS filters under the mean absolute error (MAE) and the mean squared error (MSE) criteria. Simulation results in image processing demonstrate that GWOS filters, like generalized stack filters, can suppress both impulsive noise and Gaussian noise more effectively than standard stack filters.

Design of optimal generalized stack filters (GSFs) under the mean absolute error (MAE) criterion suffers from two bottlenecks, that is, the design procedure depends on the joint statistics of the signal and noise processes that are rarely known and calls for a huge linear program (LP). In this paper we suggest efficient approaches to solve these problems. First, we present a method of estimate, based on training sequences, all the probabilities needed during the filter design procedure. Then, we introduce an algorithm that only involves data comparisons exclusively, but results in optimal filters in most practical cases. Design examples for image restoration from impulsive noise are provided.

In this paper, we present characteristic pattern as means to match shapes. The characteristic pattern is a group of convex sets that can uniquely describe the shape with respect to a pre- selected set of basic patterns or structuring elements. Each convex subset of a given shape is a union of convex sets each derived from consecutive dilation of one of structuring elements. A shape can be distinguished from others by inspecting the non-zero measure on the resultant sets of openings with a set of pre-selected elements: structuring elements and the numbers of consecutive openings. A transformation from weakly-connected sets to strongly-connected sets is introduced.

Numerous successful methodologies have been developed for gray scale image processing and analysis. However, their applications to color images are plagued because of the multi- dimensionality and intercorrelation of color images. In this paper, we examine the decorrelation of color images using the principal component analysis. Both global and local analyses are studied. To more efficiently decorrelate the images based on local features, a total color difference measurement is utilized to find the regions in which the decorrelation takes place. We call this dynamic decorrelation. Experiments are enclosed and comparisons of dynamic decorrelation with both global and local decorrelation are conducted.

Hybrid bipixel structuring element decomposition can be accomplished if the decomposed primary components of a structuring element are either line segments or parallelograms, since both have optimal bipixel decomposition. The proposed method is used for the decomposition of digital disks and can be extended to arbitrary structuring elements. Using the decomposed digital disks, Euclidean morphological dilation and erosion can be performed with linear computational complexity.

The research reported in this paper is aimed at locating, identifying, and quantifying internal (anatomical or physiological) structures, by 3-D image segmentation. Computerized tomography (CT) images of an object are first processed on a slice-by-slice basis, generating a stack of image slices that have been smoothed and pre-segmented. The image smoothing operation is executed by a spatially adaptive filter, and the 2-D pre-segmentation is achieved by a thresholding process whereby each individual pixel in the input image space is consistently assigned a label, according to its CT number, i.e., the gray-level value. Given a sequence of pre-segmented images as 3-D input scene (a stack of image slices), the spatial connectivity that exists among neighboring image pixels is utilized in a volume growing process which generates a number of well-defined volumetric regions or image solides, each representing an individual anatomical or physiological structure in the input scene. The 3-D segmentation is implemented using a volume growing process so that the aspect of pixel spatial connectivity is incorporated into the image segmentation procedure. To initialize the volume growing process for each volumetric region in the input 3-D scene, a seed location for a region is defined and loaded into a queue data structure called seed queue. The volume growing process consists of a set of procedures that perform different operations on the volumetric data of a CT image sequence. Examples of experiment of the described system with CT image data of several hardwood logs are given to demonstrate usefulness and flexibility of this approach. This allows solutions to industrial web inspection, as well as to several problems in medical image analysis where low-level image segmentation plays an important role toward successful image interpretation tasks.

In this paper a novel regular array architecture for the 2-D running order statistics sort problem will be presented. The architecture combines a high throughput in terms of sorted windows with low hardware costs and low I/O bandwidth. It will be shown that the throughput-hardware cost ratio is substantially better than that of previously published architectures for this sorting problem under the same I/O constraints. This has been achieved by a careful design of the sorting algorithm that is tuned to match as tightly as possible the needs of real-life applications requiring this type of sorting. This design illustrates that designing algorithms with both precise applications and an architectural style in mind can greatly reduce the cost of VLSI Application Specific Integrated Circuit (ASIC) implementation. In this way it enhances the feasibility of a VLSI implementation of an algorithm.

An algorithm named Minimum Cross-Entropy Algorithm (MCEA) for image reconstruction from incomplete projection data is presented here. The principle of maximum entropy has been applied to image reconstruction successfully. However, the application of principle of minimum cross-entropy to image reconstruction from incomplete projection data has not been found in literature. When the missing data or missing angle is large, the proposed algorithm yields acceptable results. Compared with the maximum entropy algorithm MENT, we conclude that (1) MCEA is superior to MENT for such cases where the number of projections involved are small; (2) Convergence performance of MCEA is better than MENT; (3) MCEA and MENT both are stable against noise; and (4) Under appropriate a priori distribution, MCEA yields satisfactory results after a couple of iterations. The speed and quality of reconstruction as well as the overhead of storage for MCEA are superior to that of MENT.

Methods for thinning connected components of an image differ in the size of support, type of connectivity preserved, degrees of parallelism and pipelining, and smoothness and fidelity to structure of the results. A unifying framework is presented, using image morphology, of all 4- and 8-connectivity-preserving (CP) transformations that use a 3 X 3 basis of support on binary images discretized on a square lattice. Two types of atomic CP transformations are defined: weak CP neither breaks nor joins components and strong CP additionally preserves the number of connected components. It is shown that out of thousands of possible 3 X 3 hit-miss structuring elements (SEs), in their most general form there are only four SEs (and their rotational isomorphs) for each of the two sets (4- and 8-connectivity) that satisfy strong CP for atomic operations. Simple symmetry properties exist between elements of each set, and duality relations exist between these sets of SEs under reversal of foreground/background and thinning/thickening operations. The atomic morphological operations, that use one SE, are intrinsically parallel and translationally invariant, and the best thinned skeletons are produced by sequences of operations that use multiple SEs in parallel. A subset of SEs that preserve both 4- and 8-connectivity and have a high degree of symmetry can be used in the most parallel fashion without breaking connectivity and produce very smooth skeletons. For thickening operations, foreground components either self-limit on convex or expand indefinitely. The self-limited convex hulls are formed either by horizontal and vertical lines, or by lines of slope +/- 1. Four types of boundary contours can result for thickening operations that expand indefinitely. Thickened text images result in a variety of typographically interesting forms.

Total divergence has been shown to be a useful measure of performance of visual recognition processes. Measurement and comparison of divergences using samples require statistical processing and sample size becomes an important factor in practical applications when sample sizes are small. We recently completed a numerical study on the applicability of total divergence to a medical Pap test process in small samples. This study suggested that in order to keep measurement biases within +/- 0.5 decits, 100 - 200 samples of cytology- histology pairs were required in the best classifications of 3, 4, and 5 category-states. At these sample sizes, measurement errors (standard errors) were also contained within +/- 0.5 decits. This study also confirmed that previously reported, over-estimated propagated errors in small samples were in fact over-estimation, and that their use for testing a null hypothesis was valid. The number of samples with indefinable statistics due to a zero denominator can be as high as 30% when the sample sizes were 500 for 3, 4, and 5 category-state classifications. Biases due to small samples were positive for most category-states except for the optimum 3 category-states, in which bias changed to negative (bias inversion), and observed errors of total divergence paradoxically decreased as N decreased (error-sample paradox) for a small sample size (N < 700).

A group theoretic approach to image representation and analysis by means of wavelet-type transforms is presented. The following special cases of interest in image representation and in biological and computer vision are discussed: 2- and 3-D rigid motion; similarities; and 2-D projective group obtained from 3-D camera rotation followed by the group of planar translation.

A new concept of Neural Network-Aided Design (NN-AD) is presented. It is a hierarchical approach consisting of several concatenated stages of visual information processing that are designed by training neural networks (NN). Thus, NN-AD can be viewed as a general tool for the design of special filters in accordance with the specific task of image processing under consideration. The nonlinear filters are formatted by a supervised presentation of a proper set of input-output patterns. The principles of NN-AD design are illustrated by the examples of edge detection with subpixel resolution and of orientational processing for edge enhancement. The proposed NN-AD approach is found to be very robust with regard to various types of errors.

Orthonormal bases of wavelets are a powerful tool in image representation and processing. A tree algorithm makes the process simple and computationally efficient. A comparative study of compactly supported wavelets shows that, when applied to the analysis of digital images, the simplest choice of Haar's wavelet is also the optimal choice.