One major challenge in contemporary networks is that energy constitutes a limited and yet a crucial resource. In this paper, we develop an energy-efficient power control game for the multi-carrier (MC) direct sequence (DS) code division multiple access (CDMA) system. Here, we pursue an efficient approach to power control in order to achieve high throughput at a cost of minimal energy consumed. Besides evaluating the Nash equilibrium in MC-DS/CDMA, we develop a new pricing strategy, which is shown to outperform previous pricing strategies without tuning or training procedure to avoid large computational complexity and processing delay. Finally, the simulation results have shown that the proposed approach achieves a significant improvement in quality of service (QoS) specified as the total system utility, and compared to the Nash equilibrium, the present algorithm is useful to combat multiple access interference (MAI) and AWGN noise power to maintain the satisfying performance.

Multiple-input-multiple-output orthogonal frequency division multiplexing (MIMO-OFDM) has been receiving a great deal of attention as a solution for high-quality service for next generation wireless communications. However one of main problems of orthogonal frequency division multiplexing (OFDM) is its high peak-to-average power ratio (PAPR) which seriously limits power efficiency of high power amplifier (HPA). In this paper, we present PAPR reduction technique of V-BLAST based MIMO-OFDM system. We use the selected mapping (SLM) technique as a PAPR reduction technique since it does not cause any signal distortion. As a special protection for side information (SI) of SLM technique, we propose SI power allocation technique. Simulation results show that proposed technique gives significantly better BER performance than ordinary SLM technique for MIMO-OFDM systems.

Energy-efficiency in wireless sensor networks is a very critical design issue, because they usually operate with tiny batteries and take a long time for replacement. In this context, we analyze the optimized scheduling strategy to minimize the energy consumed by data fusion in wireless sensor network. For implementation purpose, a low-complexity fractional-inverse- log scheduling (FILS) algorithm is presented to reduce extra significant amount of energy consumption compared to previously designed protocols. Next, in order to eliminate the communication overhead in centralized scheduling protocols, the simplified distributed fractional-inverse-log scheduling is also provided, which is shown to be very efficient in energy saving especially for a large-scale wireless sensor network. With consideration of the peak power constraint in real circuit design, we update FILS to further constrict the transmission time. Simulation results show that with peak transmission power limitation, energy consumption is still substantially reduced by FILS, and it yields more energy saving for the system with high peak-to-average power ratio (PAPR).

A computer aided detection and diagnostic system has been developed for diabetic retinopathy (DR). The system detects the fovea, blood vessel network, optic disk, as well as bright and dark lesions associated with DR. The diagnosis is based on the number, type and location of abnormalities relative to the fovea. Detection of normal retinal components was done as part of the overall system development and the work has been reported in literature. Lesion detection is accomplished through the process of eliminating the normal retinal components: blood vessels, fovea and optic disk. Remaining objects in the retinal image include the background and abnormalities if present. The image is partitioned in two regions: fovea and non-fovea, which have different backgrounds. Filtering and statistical adaptive thresholding are applied throughout the remaining data. The diagnostics and final system's layer is a knowledge-based system.

A parallel algorithm to detect retinal blood vessels has been developed for use in an automated diabetic retinopathy detection system. Localized adaptive thresholding and a multi-window Radon transform (RT) are utilized to detect the vascular system in retinal images. Multi-window parameter transforms are intrinsically parallel and offer increased performance over conventional transforms. The image is adoptively thresholded and then the multi-window RT is applied at different window sizes or partition levels. Results from each partition level are combined and morphologically processed to improve final performance. Multiple partitions are necessary to account for the size variation present in retinal blood vessels. The algorithm was tested with 20 images, 10 normal and 10 abnormal and the results demonstrate the robustness of the algorithm in the presence of noise. An average true positive rate of 86.3 % with a false positive rate of 3.9% is accomplished with this algorithm when tested against hand-labeled data

Orthogonal Frequency Division Multiplexing (OFDM) has several attributes which make it a preferred modulation scheme for high speed wireless communications. However, its high Peak-to-Average-Power Ratio (PAPR) causes nonlinear distortion thus limiting the efficiency of the transmitter's High Power Amplifier. Among the approaches proposed for PAPR mitigation, the Partial Transmit Sequence (PTS) technique is very promising since it does not generate any signal distortion. However, its high complexity makes it difficult for use in high speed communication systems. We present a new lowcomplexity tree algorithm to implement the PTS approach, which seeks the best trade-off between performance and complexity. Simulation results show that this new technique's performance is similar to that of the optimum case, however with significantly lower complexity.

Contrast is a measure of the variation in intensity or gray value in a specified region of an image. The region can be most or all of the image, giving rise to a global concept of contrast. The region might, on the other hand, be a small window in which case the concept of contrast is a locally defined expression. In this work, we introduce a nonlinear local contrast enhancement method. This method utilizes the Munsell value scale which is based upon human visual perception. Use of the Munsell value scale allows for the partitioning of the gray scale into ten discrete subintervals. A contrast enhancement function then thresholds the gray values in a subinterval in a smooth manner about a locally computed quantity called the mean edge gray value. By enhancing the contrast in this way the original shades of gray are preserved tuned to human perception

Contrast is a measure of the variation in intensity or gray value in a specified region of an image. The region can be most or all of the image, giving rise to a global concept of contrast. The region might, on the other hand, be a small window in which case the concept of contrast is a locally defined expression. In this work, we introduce a nonlinear local contrast enhancement method. This method utilizes the Munsell value scale which is based upon human visual perception. Use of the Munsell value scale allows for the partitioning of the gray scale into ten discrete subintervals. Subsequent local processing occurs within each of these subintervals. Inside each subinterval, this method constructs a contrast enhancement function that is a smooth approximation to the threshold step function and which maps a given subinterval into itself. This function then thresholds the gray values in a subinterval in a smooth manner about a locally computed quantity called the mean edge gray value. By enhancing the contrast in this way, the original shades of gray are preserved. That is, the groupings of the gray values by subinterval are preserved. As a result, no gray value distortion is introduced into the image.

The space F(H), or simply F, is a reproducing kernel Hilbert space (RKHS) of analytic (nonlinear) functional (Volterra functional) on a separable Hilbert space H. It was introduced in the late 1970's by the author, in collaboration with T.A.W. Dwyer, III, and L. Zyla, to represent input-output maps of large scale nonlinear dynamical systems. In the present paper we show how the properties of F, and, in particular, its reproducing kernel, can be used to model the structure and behavior of recurrent neural networks (RNNs).

The document was not made available for publication as part of the conference proceedings.

A multilayer perceptron (MLP) can be used to implement a vector quantizer (VQ) under severe constraints in the computational complexity allowed. Such constraints are typical in applications such as focal-plane image compression, in which we are interested in eliminating the analog-to-digital (A/D) converters and mapping the analog data directly into a compressed bit stream, to save energy and silicon area. We compare a nonlinear MLP called the kernel lattice vector quantizer (KLVQ) and a clustering MLP known as the cluster-detection-and-labeling (CDL) network, with regard to their hardware requirements. We show that for similar rate-distortion performances, the KLVQ has complexity smaller than that of the CDL network.

Alpha-trimmed mean filters are widely used for the restoration of signals and images corrupted by additive non-Gaussian noise. They are especially preferred if the underlying noise deviates from Gaussian with the impulsive noise components. The key design issue of these filters is to select its only parameter, alpha, optimally for a given noise type. In image restoration, adaptive filters utilize the flexibility of selecting alpha according to some local noise statistics. In the present paper, we first review the existing adaptive alpha-trimmed mean filter schemes. We then analyze the performance of these filters when the underlying noise distribution deviates from the Gaussian and does not satisfy the assumptions such as symmetry. Specifically, the clipping effect and the mixed noise cases are analyzed. We also present a new adaptive alpha-trimmed filter implementation that detects the nonsymmetry points locally and applies alpha-trimmed mean filter that trims out the outlier pixels such as edges or impulsive noise according to this local decision. Comparisons of the speed and filtering performances under deviations from symmetry and Gaussian assumptions show that the proposed filter is a very good alternative to the existing schemes.

The use of multiple antennas at the transmitter and receiver results in enormous capacity increase. However, the performance of many current space-time architectures like Vertical Bell Laboratories Layered Space-Time (V-BLAST) are not optimum in the maximum likelihood (ML) sense. We present a new algorithm, which aims at bridging this gap for the V-BLAST system. Both optimum detection and original V-BLAST detection are special cases of this algorithm. In addition, any in-between performance-complexity can be achieved by suitably choosing two parameters. Moreover, with little increase in computational power, a very high improvement in error performance over the original V-BLAST algorithm is achieved.

This paper presents a novel algorithm for edge detection from an image that is corrupted by additive Gaussian noise. The proposed approach uses a criterion for finding pixels where the image gradient is dominated by the noise term. From these pixels, it produces an approximate distribution of the noise gradient magnitude, and then, using this distribution, a threshold for a given false alarm rate is obtained. After grouping the pixels into structured clusters, the cluster average is tested against the threshold to detect the edge. The performance of the proposed method was found to be superior to other automatic thresholding methods tested.

This paper further explores the concept of line-support regions for extracting edges from low contrast images. The current method obtains the support regions by using clustering-based segmentation and refines the support regions by using split and merge techniques. Alternative methods for fitting line to cluster are presented.

This paper presents a number of concepts that are important contributions to the area of contrast enhancement. Among them are the idea of gray scale partitioning, the use of a tunable cubic polynomial for the contrast enhancement function, and, in the case of noisy images, the use of a tandem of pyramidal lowpass filters to remove the noise from homogeneous regions (regions of constant or near-constant intensity) while simultaneously preserving the edges. Gray scale partitioning is a unique and original concept. This idea is based upon the human perception of a set of gray chips. The human perception was quantified by a set of ten intensity values that were converted to reflectances. The essence of gray scale partitioning is to transform these reflectances into gray scale intensity values (assuming constant illumination). These intensity values form the endpoints of discrete subintervals and are therefore the basis for gray scale partitioning. This ensures that the enhancement process will preserve shades of gray. That is, the output gray value will be of the same shade of gray (will lie in the same subinterval) as the input gray value.

We study time series prediction capabilities of a new type of artificial neural networks, called dynamical-functional artificial neural networks (D-FANNs). These are two-layer neural systems in which the synaptic weights are "functions" rather than numbers, and where the action of a synapse on a signal passing through it takes place in the form of a scalar product in L/sup 2/ between the functional weight and the signal. The functional weights of the first layer of a D-FANN are modeled as the impulse responses of a set of linear time-invariant differential dynamical systems. We consider only the special discrete case when these impulse responses are the set of discrete signals corresponding to the discrete cosine transform. The time series prediction performance of this D-FANN model is demonstrated on a speech signal with high dynamics.

In this paper we introduce a new design strategy for SFWO (sampled function weighted order) filters based on approximation theory. It is shown that with a good choice of noise tail-length estimator and a good approximation of the relation between this estimate and filter parameters, the adaptive SFWO filter gives very promising results when the noise type is unknown or varies with time. For this use, we also introduced a new tail-length estimator based on robust statistics.

We propose in this paper a new unsupervised neural network which is capable of clustering a set of experimental data according to a given generic interpoint similarity measure, and then assign to each new input its appropriate cluster label. The network is able to do this for clusters of any shape, and without knowing in advance the number of clusters to be created. We call this new two-layer network a cluster-detection-and-labeling (CDL) network. In the CDL network the concept of similarity and closeness with regard to distance are combined. Specifically, clusters are represented by a set of prototypes, and the similarities between an input vector and these prototypes are calculated as the inner products of these vectors compared to some thresholds. These thresholds, which depend on the distance between the input vector and the prototype, are calculated in a separate threshold calculating unit. During clustering, the data are cycled through the network several times. At the end of each cycle the clusters are evaluated, and only those with more than a specified number of samples are retained. The others are fed back to be reclustered by an updated network. This process terminates according to a suitable criterion, such as when a prespecified portion of the data are classified. The performance of the CDL network has been compared with that of the winner-take-all (WTA) network for several different cluster structures, since the latter is widely used in cluster analysis applications. These studies demonstrate that the new network performs well for all the tested cluster shapes, also for those cases where the WTA network completely fails.

This paper presents an object-based video modeling. Motion segmentation is performed at the initial frame to identify different coherently moving regions, called motion-objects. These regions are grouped to form objects. Each motion-object is fitted a content-based mesh, and tracked subsequently to the next frame via mesh motion estimation and compensation. The uncovered background region(s), which emerges when objects move, will be segmented so as to identify the new objects or occluded parts of already existing objects. The mesh model is modified to reflect the changes in object boundaries.

This paper describes learning rules using simultaneous perturbation for a neurocontroller that controls an unknown plant. When we apply a direct control scheme by a neural network, the neural network must learn an inverse system of the unknown plant. In this case, we must know the sensitivity function of the plant using a kind of the gradient method as a learning rule of the neural network. On the other hand, the learning rules described here do not require information about the sensitivity function. Some numerical simulations of a two-link planar arm and a tracking problem for a nonlinear dynamic plant are shown.

This paper presents a preliminary study of a new class of nonlinear (Volterra) filters which reduce the noise in an image while simultaneously enhancing the contrast. The design of these filters draws heavily from the theory of generalized Fock (GF) spaces of Volterra series. The structure of these filters is obtained by an orthogonal projection in a GF space under the data constraints. As a consequence, even though this structure embodies an infinite Volterra series, it is represented in a closed form as a linear combination of exponentials, the exponents of which are linear functions of the input image vector.

We present a methodology for scene analysis using contextual feedback and demonstrate the approach in a simple example. In our approach, we combine bottom-up and top-down decision making into a systematic mechanism. From feedback of high-level descriptions, the system iteratively arrives at improved identification. The implementation uses OINNs (optimal interpolative neural networks) to determine plausibility and enforce constraints in label assignments. The proposed method can be extended for the analysis of more complex scenes.

An adaptive order filter structure for image restoration, which utilizes a specific class of filters commonly known as L-filters is presented. The coefficients of the present filter change locally depending on the value of the robust estimate of the local standard deviation. The form used for this estimate is the median of the absolute deviations from the median (MAD). The proposed filter can adapt well to a variety of noise types. Computer simulations illustrate the new filter to be very competitive in restoration of images corrupted by different noise types without any prior knowledge about the noise.

The conventional image intensity histogram has played a major role in some image processing operations, such as image enhancement. In this paper, we introduce two related concepts. First, we define the gradient histogram of an image as an entity which describes the (statistical) orientation properties of the image. Using gradient histograms, we then introduce a class of image orientation signatures, periodic functions of one variable which characterize the image and possess useful translation and scaling invariance properties. Orientation signatures have been successfully applied to the problem of 3D object attitude determination from a single camera image.

In this paper, illumination control has been studied as a means of enhancing image features. Such features are points, edges, and shading patterns, which provide important cues for the interpretation of an image of a scene and the recognition of objects present in it. Based on approximate expressions for the reflectance map of Lambertian and general surfaces, a rigorous discussion on how intensity gradients are dependent on the direction of the light is presented. Subsequently, three criteria for the illumination of convex-shaped cylindrical surfaces are given. Two of these, the contrast equalization criterion and the max-min equalization criterion, are developed for optimal illumination of convex polyhedrons. The third, denoted shading enhancement, is applicable for the illumination of convex curved objects. Examples illustrate the merit of the criteria presented.

Kinematic redundancy of robotic manipulators has been known to be useful in accomplishing additional performance on top of the main task of tracking desired end effector trajectories. Robustness has been a major concern in manipulator control where exact physical properties of manipulators are unobtainable. In this paper, we review a method of analyzing robustness of joint space controllers for manipulators, and extend the analysis to end effector space control for the first time. We then generalize the approach to the case of kinematically redundant manipulators by introducing the redundancy as part of the manipulator states, and show that kinematic redundancy is useful in making manipulator controllers more robust. The final result of this paper is an optimization problem whose solution determines the self-motion which optimizes the system robustness.>

This paper provides a brief overview of a rigorous framework, developed by the author, for the modeling and identification of nonlinear dynamical systems by artificial neural networks. The system model is obtained as a best approximation of the operator(s) representing the system in a "neural space", under interpolating or smoothing constraints imposed by the input-output training data. This optimal modeling results in one of four types of neural networks proposed and discussed by the author elsewhere, namely the OI, OS, OMNI and OSMAN nets. The identification of a system so modeled can take place instantaneously by batch processing of the training data, or sequentially by adaptation, learning, and/or evolution.>

A discussion of the step edge model and the zero crossings of edge detectors is provided to justify the authors' previous analysis (see ibid., vol. 12, p. 1186-1189) of the localization performance error. It is shown that the previous analysis is consistent with its assumptions. The exact solution to the zero crossing density of the edge detection filter is also provided.

The unified approach to optimal image interpolation problems presented provides a constructive procedure for finding explicit and closed-form optimal solutions to image interpolation problems when the type of interpolation can be either spatial or temporal-spatial. The unknown image is reconstructed from a finite set of sampled data in such a way that a mean-square error is minimized by first expressing the solution in terms of the reproducing kernel of a related Hilbert space, and then constructing this kernel using the fundamental solution of an induced linear partial differential equation, or the Green's function of the corresponding self-adjoint operator. It is proved that in most cases, closed-form fundamental solutions (or Green's functions) for the corresponding linear partial differential operators can be found in the general image reconstruction problem described by a first- or second-order linear partial differential operator. An efficient method for obtaining the corresponding closed-form fundamental solutions (or Green's functions) of the operators is presented. A computer simulation demonstrates the reconstruction procedure.

The calculation of the innovation process plays a significant role in the solution of the nonlinear filtering problem. Convergence of an approximate calculation of the innovation process is proved, and then illustrated.>

An evolution-oriented learning algorithm is presented for the optimal interpolative (OI) artificial neural net proposed by R. J. P. deFigueiredo (1990). The algorithm is based on a recursive least squares training procedure. One of its key attributes is that it incorporates in the structure of the net the smallest number of prototypes from the training set T necessary to correctly classify all the members of T. Thus, the net grows only to the degree of complexity that it needs in order to solve a given classification problem. It is shown how this approach avoids some of the difficulties posed by the backpropagation algorithm because of the latter's inflexible network architecture. The performance of this new algorithm is demonstrated by experiments with real data, and comparisons with other methods are also presented.

A generalization of the concepts of moments and moment invariants to n-dimensions is presented. These concepts are used to develop techniques for object identification and attitude determination. Experimental results for these techniques are presented, and theoretical error analyses are developed.>

It is shown how choosing the self-motion inherent in redundant arms is crucially important when flexibility is present. The self-motion, by exciting or damping the flexural modes, alters the dynamic response of the arms. Kinematic redundancy can also be used in many cases to help damp out vibrations. This issue is examined, and control algorithms designed to regulate the flexibility while maintaining precise tracking of the end-effector trajectory are introduced. The controllers are of the computed torque type in end-effector space, and use self-motion of the links to reduce the flexible effects.>

A two-stage design method for artificial neural networks is presented. The first stage is an evolutionary RLS (recursive least squares) algorithm which determines the optimal configuration of the net based on the concept of optimal interpolation. During this stage, the members of a given sample set are processed sequentially and a small consistent subset, constituting what we call prototypes, is selected to form the building blocks of the net. The synaptic weights as well as the internal dimensions of the net are updated recursively as each new prototype is selected. The evolving net at each intermediate step is a modified version of the Optimal Interpolative (OI) net derived in a recent paper by one of the authors. The concept of an evolving network configuration is attractive since it does not require the prescription of a fixed configuration in order to learn the optimal synaptic weights. This can eventually lead to a network architecture which is only as complex as it needs to achieve a required interpolation function. The second stage is for the fine adjustment of the synaptic weights of the network structure acquired during the first stage. This stage is a two-step iterative optimization procedure using the method of steepest descent. The initial values of the synaptic weights in the iterative search are obtained from the first stage. It is seen that they are indeed very close to the optimal values. Hence, fast convergence during the second stage is guaranteed.

A theory of photometric stereo is proposed for a large class of non-Lambertian reflectance maps. The authors review the different reflectance maps proposed in the literature for modeling reflection from real-world surfaces. From this, they obtain a mathematical class of reflectance maps to which the maps belong. They show that three lights can be sufficient for a unique inversion of the photometric stereo equation for the entire class of reflectance maps. They also obtain a constraint on the positions of light sources for obtaining this solution. They investigate the sufficiency of three light sources to estimate the surface normal and the illuminant strength. The issue of completeness of reconstruction is addressed. They show that if k lights are sufficient for a unique inversion, 2k lights are necessary for a complete inversion.>

Consideration is given to an optimal modeling problem for the general horizontally stratified ocean, which is assumed to be a random inhomogeneous medium. The random acoustic wave velocity potential p=p(x, y, z) in the ocean is described by the stochastic Helmholtz equation Delta p+k/sup 2/p=0, where Delta is the standard Laplacian partial differential operator, and k=k(x, y, z) is the random wave number. Because of the random and inhomogeneous nature of the ocean, not only the amplitude and phase fluctuations of the sound wave but also its scattering have to be taken into account in the mathematical modeling of the random wave propagation. In this consideration, p is decomposed as p=p/sub 0/+p/sub s/, where p/sub 0/ is the deterministic component (the unperturbed part) of the velocity potential (which can be determined by some existing techniques) and p/sub s/ is the random component created essentially by a perturbation. A novel method for obtaining an optimal estimate of p/sub s/ is proposed. It is based on a PDL/sub g/-spline theory and technique developed by the authors, so that p can be determined optimally and efficiently for the purpose of applications.>

The use of structured interconnection topologies is considered for designing layered feedforward neural networks when no explicit information on the application data space is available at design time. Sparse, structured designs can yield computationally efficient networks when the input and output data spaces are large. In particular, the fully connected, cube connected, and butterfly connected networks are analyzed, and their interconnection characteristics are compared. The tradeoff incurred in making the incidence matrix increasingly more sparse is noted, and it is concluded that an acceptable design should be moderately interconnected, allowing a reduction in synaptic density while maintaining functional flexibility.>

The authors address the issue of whether the shape of an object and certain parameters of its reflectance map can be jointly estimated using photometric stereo. This problem has been addressed in the literature for the case where the Lambertian and non-Lambertian components in the image can be easily separated. The authors do not assume each separability. They highlight some aspects of the detailed solution to the above problem.>

A novel optimal model is obtained for the matching score subnet of a pattern classifier network by requiring that the model interpolate a set of exemplary input-output pairs while optimizing a min-max criterion in an uncertain set Γ a neural spaceFto which the net is assumed to belong. This criterion is presumed to provide robustness to the optimal model obtained. The optimal model consists of a two-layer feedforward net, the first layer being the same as the matching score layer of the Hamming net, and the second layer being a new layer (absent in the Hamming net) possessing synaptic weights which result from the optimization procedure used. Applications of this net to some typical signal- and image-classification problems are being pursued

The localization performance measure of edge detectors is addressed. A one-dimensional formulation of the problem is considered. A linear space-invariant filter is used for the detection. The locations of local maxima in the thresholded output of the filter are declared to be the edges. The limitations of conventional performance measures are shown, and a localization performance measure for edge detection is suggested. This performance measure is based on the theory of zero-crossings of stochastic processes. It is shown that the derivative of a Gaussian is the optimal edge detector for the measure.>

The authors present the main results of a theory of photometric stereo for non-Lambertian surfaces. First, they propose a class of reflectance maps that can be used for a realistic representation of real-world reflection. Second, they ask how many light sources are needed for inverting this class of reflectance maps. It is shown that three light sources are sufficient for a globally unique inversion for the entire class of reflectance maps. Finally, the authors address the issue of completeness of reconstruction. They show that if k lights are sufficient for a globally unique inversion, 2k lights are necessary for a complete inversion.>

The problem of model-based optimal reconstruction of an image from its samples is studied. It is assumed that the image signal is described by a linear PDE (partial differential equation) model from which a sequence of output frames is available. These frames need not have rectangular boundaries or be uniformly sampled. Under the criterion that the image to be reconstructed is the one that is created by an unknown input signal with minimum energy, the authors obtain a unique optimal solution by interpolating the image samples with a recently formulated generalized spline called the PDL/sub g/ spline. It is possible to show that such a reconstruction corresponds to a minimum-mean-squares estimate of the image, given its samples. The authors also derive an algorithm for finding the optimal solution in an explicit and closed form. They include some computer simulation results to show the quality of the images obtained with the method.>

A procedure is described for translating high-level knowledge about an application into a neural network architecture that incorporates that knowledge in its synaptic interconnections. The program's design philosophy stresses separation of neural design from network function, a uniform syntax for neurons, inputs, and outputs, and flexibility in modularizing the resulting network. The trainability of neural networks allows the application knowledge to be further fine-tuned for a specific data space. It is shown how the translation rules can allow for the selective training of subnetworks, and an example of a signal recognition algorithm to which this technique can be applied is given.>

In filter design theory, one derives the filter structure by requiring that the frequency response of the filter match, in some optimal way, a desired frequency response. A similar philosophy is adopted to derive the architecture of a neural net, except that in this case, because of the nonlinear nature of the neural net, the frequency response can no longer be used as a design criterion. Instead, it is required that the net N, whose structure is to be determined, match a set of exemplary input-output pairs while minimizing the maximum error in some appropriate space. The latter criterion guarantees the robustness of the net.>

The authors present a constraint formulation that yields a real-numbered solution to the frequency-domain Prony algorithm. This is done by requiring that all the roots of the Z-transform of the solution be on the unit circle. They make appropriate transformations that map this unit circle onto a left half plane, which then allows them to use Sturn's theorem to check the number of roots in the left half plane. This newly formulated constraint is incorporated in the Prony method in the frequency domain. One of the applications of the method is in the decomposition of a mixture of delayed versions of a signal, where the constraints guarantee that the solution for the time delays are real.>

Returned radar power estimates are used in an iterative procedure which generates successive approximations to the target shape in order to determine the shape of a 3-D surface. A simulation is shown which involves the reconstruction of an edge of a flat plate. Although this is a somewhat artificial example, it addresses the real problem of recovering edges of space objects lost in shadow or against a dark background. The results indicate that a microwave/optical sensor fusion system is possible, given sufficient computing power and accurate radar cross section measuring systems.>

The Darboux vector contains both the curvature and torsion of a three-dimensional (3-D) image as its components. Curvature measures how sharply a curve is turning while torsion measures the extent of its twist in 3-D space. Curvature and torsion completely define the shape of a 3-D curve. A scheme is presented that uses the length of this vector, also called the total curvature, for the segmentation of 3-D contours. A quintic B-spline is used in this formulation to obtain the total curvature for noisy data. Examples of nonnoisy and noisy data illustrate the merit of the scheme.>

In this paper, we investigate in detail the properties of the maximum likelihood estimator of the generalized p-Gaussian (gpG) probability density function (pdf) from N independent identically distributed (iid) samples, especially in the context of the deconvolution problem under gpG white noise. In the first part, we describe the properties of the estimator independently of the application. In the second part, we obtain the solution of the above-mentioned deconvolution problem as the solution of a minimum norm problem in an lpnormed space. In the present paper, we show that such a minimum norm solution is the maximum likelihood estimate of the system function parameters and show that such an estimate is unbiased, with the lower bound of the variance of the error equal to the Cramer-Rao lower bound, and the upper bound derived from the concept of a generalized inverse, both of which we also give.

Waveforms may be represented symbolically such that their underlying, global structural composition is emphasized. One such symbolic representation is the relational tree. The relational tree is a computer data structure that describes the relative size and placement of peaks and valleys in a waveform. Researchers have developed various distance measures which serve as tree metrics. A tree metric defines a tree space. We are able to cluster groups of trees by their proximity in a tree space. Linear discriminants are used to reduce vector space dimensionality and to improve cluster performance. A tree transformation operating on a regular tree language accomplishes this same goal in a tree space. Under certain restrictions, relational trees form a regular tree language. Combining these concepts yields a waveform recognition system. This system recognizes waveforms even when they have undergone a monotonic transformation of the time axis. The system performs well with high signal to noise ratios, but further refinements are necessary for a working waveform interpretation system.

In the past decade, vision/sensing systems techniques have received significant attention for applications in robotics and automation. The thrust of automation and robotics for space applications has been proposed for increased productivity, improved reliability, increased flexibility, higher safety, and for the performance of tasks unsuited to humans. These benefits can be insured through automating time-consuming tasks, increasing productivity/performance of crew-accomplished tasks, and performing tasks beyond the capability of the crew. This paper provides a review of efforts currently in progress at the NASA/Johnson Space Center and at Rice University, the accomplishments to date, and some of the anticipated future developments. Both systems and algorithms are discussed. The evolution of future vision/sensing is projected to include the fusion of multisensors ranging from microwave to optical with multimode capability to include position, attitude, recognition, and motion parameters. The algorithms for information extraction would incorporate aspects of intelligence and knowledge for the interpolation and extrapolation of the needed data. The key features of the overall system design will be small size and weight, fast signal processing, robust algorithms, and accurate parameter determination. These aspects of vision sensing will also be discussed in this paper.

This paper presents a new scheme for segmentation of three-dimensional (3D) contours based on a criterion involving their curvature and torsion properties. Curvature indicates how sharply a curve is turning while torsion measures the extent of its twist in 3D space. Also, curvature and torsion completely define a curve in 3D space. For these reasons, the two together have been chosen here as a set of intrinsic and complete structural features of a contour. The Darboux vector Ω contains both these features as its components. We use the length ∥Ω∥ of the Darboux vector, also known as total curvature, as the criterion for segmentation of the contour. Specifically we choose the zero-crossings of d2∥Ω∥/ds2, where s = arc length, as the breakpoints in the segmentation of the contour. The segmentation scheme is described by means of a block diagram which includes filters for calculation of d2∥Ω∥/ds2in terms of appropriate distributional derivatives resulting from the contour parametric equations.

A consistent development of general moment invariants of affine transformations for two-dimensional image functions is presented. Based on this development, a new general moment-invariants/attributed-graph (MIAG) method is presented for the identification of three-dimensional objects from a single observed image using a model-matching approach. The three-dimensional location and orientation parameters of the object are also obtained as a byproduct of the matching procedure. The scheme presented allows the observed object to be partially Occluded. For identification purposes, a three-dimensional object is represented by an attributed graph describing the geometrical structure and shape of the surface bounding the object. In such a description, two-dimensional general moment invariants of the rigid planar patches (RPP) constituting the object faces are used as attributes or feature vectors which are invariant under three-dimensional motion. With this representation, the identification problem becomes a subgraph isomorphism problem between the observed image and a library model. An algorithm is presented for this matching process, and the results are illustrated by computer simulations.

A framework and method of solution is presented for interpolation and smoothing of signals by simultaneous optimization of time- and frequency-domain criteria. The solution invariably takes form of an appropriate generalized spline.

A recently developed syntactic/semantic method for optimally reconstructing a 3D surface by triangular tiling is outlined. Results of a computer simulation are shown.

We present new algorithms for rapid identification and three-dimensional (3D) attitude determination of a solid object from a single image, using a model matching approach. Our scheme allows the observed object to be partially occluded. The object, as well as the model to which it is matched are represented by the 3D surface constituting their boundaries. We assume that these surfaces consist of flat faces, i.e. the object and the model are (not necessarily convex) polyhedra. We represent each by an attributed graph. The nodes of the graph denote faces on the surface and edges indicate the adjacency of faces. Attributes on the nodes are features invariant to 3D motion made up of 2D moment invariants. With this representation the recognition problem becomes a subgraph matching problem between the image of the observed object and the stored models. We present an algorithm for the matching process, furthermore the exact attitude is obtained as a byproduct of the matching procedure.

As a generalization of the median and rank filters [1]-[3], the concept of an order filter as a useful signal and image processing tool is introduced. This filter obtains a best estimate of the signal based on its order statistics. The optimal design of such a filter is illustrated in the context of the problem of envelope estimation. Other applications are also briefly mentioned.

As an extension of the existing theory, iterative restoration algorithms are presented in this paper for the solution of noisy, nonlinear restoration problems. The constraints and distortions are assumed to be known and are modeled as general (not necessarily linear) projections. The constraints are divided into two fundamentally different types, system constraints and signal constraints, of which only the latter are susceptible to noise. These two types of constraints are imposed in separate steps, called the constraint and replacement steps, respectively. Comparison of the signals produced at each step enables one to determine whether or not the generated sequence of signals is converging. The problem caused by additive noise is overcome by modifying the replacement step. Such a modified replacement, called generalized replacement, improves both the quality of the reconstructed or restored signal and the convergence properties of the algorithm. The two types of generalized replacements considered in detail are partial replacement and windowed replacement.

A scheme for a spatial domain image data preprocessing decimation and postprocessing interpolation is presented. The scheme is implemented by appropriate FIR digital filters. Frequency and patial domain specifications are discussed in the design of the corresponding digital filters. Fast approximation techniques in the spatial domain are presented.

The magnitude error function is the optimal criterion for detection and parameter estimation under Laplacian (double-sided exponential density) noise. Its use is compared to the commonly used squared error function (equivalent to the matched filter), derived under the condition of Gaussian (normal density) noise, and found to possess the following advantages: (1) The performance of the magnitude error detector is better than that of the squared error detector for signals with unknown delays in low-level noise and within overlappings; (2) The algorithms are simpler to implement on a digital processor and they run faster because no multiply operations are needed and because the dynamic range is suitable for a 16-bit arithmetic. The theory is developed for a detection and estimation scheme on nonlinear functions of non-random parameters. It is applied to a problem in electromyography (EMG), namely the problem of separating superpositions of finite-duration signals which overlap both in the time and frequency domains. For the sake of completeness it is shown that the amplitude density of the noise in EMG is Laplace distributed rather than Gaussian.

Image transform coding is first briefly reviewed using conventional viewpoints. Then a new spatial domain interpretation is given to image transform coding. An improvement based on this viewpoint for the Fourier transform coding, which possesses simple spatial domain relations, is presented.

A time-domain approach is presented for separating a superposition of signals which overlap both in the time and frequency domains. The method is partly based on the cross-correlations among the signals and some of their derivatives. Given the data, an initial detection scheme formulates a set of hypotheses on the sets of signals which may be present in the mixture, and provides estimates of the signal positions for each hypothesis. In the second stage, the most likely hypothesis is selected and accurate values of the signal amplitudes and their time positions are determined by a nonlinear least squares approach. The method is illustrated by a simulation with emg signals.

A technique is presented for feature extraction of a waveform y based on its Tauberian approximation, that is, on the approximation of y by a linear combination of appropriately delayed versions of a single basis function x, i.e., y(t) = ¿M i = 1 aix(t - ¿i), where the coefficients ai and the delays ¿i are adjustable parameters. Considerations in the choice or design of the basis function x are given. The parameters ai and ¿i, i=1, . . . , M, are retrieved by application of a suitably adapted version of Prony's method to the Fourier transform of the above approximation of y. A subset of the parameters ai and ¿i, i = 1, . . . , M, is used to construct the feature vector, the value of which can be used in a classification algorithm. Application of this technique to the classification of wide bandwidth radar return signatures is presented. Computer simulations proved successful and are also discussed.

A mixed combinatorial and analytic approach is presented for thematic filtering of simple patterns in a scene and for interpretation of how they are arranged in the form of a complex pattern. By "simple patterns" we mean basic objects like cars, trucks, or trees while a complex pattern could be an arrangement of such objects in an outdoor scene. In our procedure, the simple patterns of interest are identified by comparison with "reference simple patterns" (reference images) via comparison of features associated with local configurations in the test and reference images. The features are invariant with respect to translation, rotation, and scaling. The procedure also gives the relative positions and orientation of the simple patterns found and thus provides the required interpretation information.

An adaptive nonlinear Kalman-type filter is presented for the restoration of two-dimensional images degraded by general image formation system degradations and additive white noise. A vector difference equation model is used to represent the degradation process. Due to the nonstationarity of an image the object plane distribution function, i.e. the original image, is partitioned into disjoint regions based on the amount of spatial activity in the image. Difference equation models are used to characterize each of the regions of this nonstationary object plane distribution function. Features of the restoration filter include the ability to account for the response of the human visual system to additive noise in the image; a two-dimensional interpolation scheme to improve the estimates of the initial states in each region; and a nearest neighbor algorithm to choose the previous state vector for the state of pixel (i,j).

Given a signal x which is of bounded energy and either time-limited-and-essentially band-limited or (strictly) bandlimited, we present techniques, based on a generalized spline framework, for optimal determination of the values of x(t) and of its spectrum X(ω) for any specified t and ω, from a finite set of samples of X.

A technique is described for recognition of contours in a spectrogram associated with a given string of events by means of dynamic models expressing relationships between frequency as the dependent variable and time as the independent variable.