Summary form only given. High-speed network traffic prediction is an essential step in building effective preventive congestion control schemes. This paper addresses the problem of high-speed network traffic prediction at different timescales. We propose a two-stage mixture model where the first stage includes two individual models, namely k-factor Gegenbauer ARMA (GARMA) and multilayer perceptron (MLP). The k-factor GARMA captures the short- and long-range dependency, whereas the MLP captures the nonstationarity. The second stage combines the two forecasts to enhance the prediction accuracy and merge the traffic characteristics captured by the individual models. Four different combination schemes are investigated. They are averaging, Karmarker's linear programming algorithm, MLP combiner and fuzzy neural network. The performance is tested on four different real traffic data, MPEG video, JPEG video, Ethernet and Internet. The problem of one-step-ahead traffic prediction at different timescales is considered. The results indicate that the proposed two-stage mixture model outperforms the individual models. The results also show that the prediction performance depends on the traffic nature and the considered timescale.

Gegenbauer autoregressive moving average (GARMA) model has the ability to capture both the short- and long-range dependent characteristics of the underlying data. GARMA has been used for modeling and forecasting of the financial time series that exhibits a long-range dependency (LRD). Since the high-speed network traffic exhibits a high degree of LRD characteristic, GARMA could be used for its modeling and prediction. In this paper, we present a simplified parameter estimation procedure and an adaptive prediction scheme for the k-factor GARMA model. The adaptation gives the model the ability to capture, the non-stationary characteristic of the data. The k-factor GARMA is applied to model four different types of real traffic data: MPEG and JPEG video, Ethernet and Internet. These models are then used to predict one-step-ahead traffic value at different timescales. The results show that the estimated parameters of the k-factor GARMA model provide a detailed and accurate presentation for the traffic characteristics in both time and frequency domain. We also demonstrate that the prediction performance of the k-factor GARMA model outperforms that of the traditional autoregressive (AR) model.

The work presents a predictive dynamic bandwidth allocation (PDBA) scheme that updates the allocated bandwidth periodically to minimize the queue's build-up process. It requires an accurate traffic predictor so we use a two-stage predictor. The first stage includes FARIMA and FNN models running in parallel to predict the traffic data. While FARIMA captures the self-similarity, FNN captures the non-stationarity. The second stage combines the two forecasts using FNN to enhance the prediction accuracy. The performances of the PDBA scheme and the predictors are tested on MPEG and JPEG data. The results show that the two-stage predictor outperforms the individual ones. The proposed PDBA results in lower CLR compared to non-predictive schemes.

The k-factor Gegenbauer ARMA process has been a widely used tool for modeling and predicting financial time series. This model has the ability to capture both short-range and long-range dependencies (SRD and LRD, respectively). Since the high-speed network traffic exhibits both of these dependencies, GARMA could be used for its modeling and prediction. In this paper, we present a simple parameter estimation procedure for the k-factor GARMA model. We also discuss the synthesis scheme for this model. The estimation procedure is used to model an MPEG video traffic data. Several synthetic traffic data are generated by varying the GARMA model parameters. The main characteristic of the synthetic traffic, self-similarity, is tested and compared to that of the actual MPEG data and the traffic synthesized using the autoregressive (AR) model. The results show that the k-factor GARMA model captures the main traffic characteristics in both time and frequency domain. They also show that the synthetic traffic data using the k-factor GARMA model are the closest to the actual data.

Measurements of high-speed network traffic have shown that traffic data exhibits a high degree of self-similarity. Traditional traffic models such as AR and ARMA are not able to capture this long-range-dependence making them ineffective for the traffic prediction task. In this paper, we apply the fractional ARIMA (F-ARIMA) model to predict one-step-ahead traffic value at different timescales. F-ARIMA has the ability to capture both the short- and long-range dependent characteristics of the underlying data. We present a simplified adaptive prediction scheme to reduce the F-ARIMA computational complexity. The performance of the proposed F-ARIMA prediction model is tested on four different types of traffic data: MPEG and JPEG video, Ethernet and Internet. We also apply the F-ARIMA prediction model to a dynamic bandwidth allocation scheme. The results show that the performance of F-ARIMA outperforms the AR model. They also show that the prediction performance depends on the traffic nature and the timescale.

This paper presents a method for the extraction of contour lines and other geographic information from scanned color images of topographical maps. Although topographic maps are available from many suppliers, this work focuses on United States Geological Survey (USGS) maps. The extraction of contour lines, which are shown with brown color on USGS maps, is a difficult process due to aliasing and false colors induced by the scanning process and due to closely spaced and intersecting/overlapping features inherent to the map. These difficulties render simple approaches such as clustering ineffective. The proposed method overcomes these difficulties using a multistep process. First, a color key set, designed to comprehend color aliasing and false colors, is generated using an eigenvector line-fitting technique in RGB space. Next, area features, representing vegetation and bodies of water, are extracted using RGB color histogram analysis in order to simplify the next stage. Then, linear features corresponding to roads and rivers including contours, are extracted using a valley seeking algorithm operating on a transformed version of the original map. Finally, an A* search algorithm is used to link valleys together to form linear features and to close the gaps caused by intersecting features. The performance of the algorithm is tested on a number of USGS topographic map samples.

The focus of this paper is on combination of artificial neural-network (ANN) forecasters with application to the prediction of daily natural gas consumption needed by gas utilities. ANN forecasters can model the complex relationship between weather parameters and previous gas consumption with the future consumption. A two-stage system is proposed with the first stage containing two ANN forecasters, a multilayer feedforward ANN and a functional link ANN. These forecasters are initially trained with the error backpropagation algorithm, but an adaptive strategy is employed to adjust their weights during on-line forecasting. The second stage consists of a combination module to mix the two individual forecasts produced in the first stage. Eight different combination algorithms are examined, they are based on: averaging, recursive least squares, fuzzy logic, feedforward ANN, functional link ANN, temperature space approach, Karmarkar's linear programming algorithm and adaptive mixture of local experts (modular neural networks). The performance is tested on real data from six different gas utilities. The results indicate that combination strategies based on a single ANN outperform the other approaches.

The focus of this paper is on combination of artificial neural network (ANN) forecasters with application to the prediction of daily natural gas consumption needed by gas utilities. A two-stage system is proposed with the first stage containing three ANN forecasters. The first forecaster is a multilayer feedforward network trained with backpropagation, the second one is another multilayer feedforward network trained with Levenberg-Marquad algorithm, and the third one is a one-layer functional link network These three separate forecasts are nonlinearly combined in the second stage using a functional link ANN combiner. A scheme is introduced to make all of the ANNs adaptive, with their weights changing throughout the forecasting phase. The performance is tested on real data from four different gas utilities for a period of several months. The results show that the proposed forecast combination approach does result in more accurate forecasts compared to using a single forecaster. The overall performance of the system is also quite good from an operational point of view.

The electrical connections between an integrated circuit (IC) die and the package external leads are made by gold wires in a process called wire bonding. This task is done with a special wire bonding machine. This machine has several parameters to set. These settings are dependent on many factors, including wire size, die material, package size, etc. At present, the machine operator sets these parameters in an ad-hoc fashion based on guidelines and their experience. This paper describes an iterative neural network based closed loop controller for regulating two important characteristics of wire bond, ball size and shear strength density. This controller provides the settings for four parameters that have major impact on these two metrics. The performance is tested on actual machine; results show successful process control.

Considers the problem of estimating parameters of multispectral random field (RF) image models using maximum likelihood (ML) methods. For images with an assumed Gaussian distribution, analytical results are developed for multispectral simultaneous autoregressive (MSAR) and Markov random field (MMRF) models which lead to practical procedures for calculating ML estimates. Although previous work has provided least squares methods for parameter estimation, the superiority of the ML method is evidenced by experimental results provided in this work. The effectiveness of multispectral RF models using ML estimates in modeling color texture images is also demonstrated.

Random field (RF) models have widespread application in image modeling and analysis. The effectiveness of these models is largely dependent on the choice of neighbor sets, which determine the spatial interactions that are represented by the model. We consider the problem of selecting these neighbor sets for simultaneous autoregressive and Gauss-Markov random field models, based on the correlation structure of the image to be modeled. A procedure for identifying appropriate neighbor sets is proposed, and experimental results which demonstrate the viability of this method are presented.

Multispectral extensions to the traditional gray level simultaneous autoregressive (SAR) and Markov random field (MRF) models are considered. Furthermore, a new image model is proposed, the pseudo-Markov model, which retains the characteristics of the multispectral Markov model, yet admits to a simplified parameter estimation method. These models are well-suited to analysis and modeling of color images. For each model considered, procedures are developed for parameter estimation and image synthesis. Experimental results, based on known image models and natural texture samples, substantiate the validity of thee results.

The long correlation (LC) models are a general class of random field (RF) models which are able to model correlations, extending over large image regions with few model parameters. The LC models have seen limited use, due to lack of an effective method for estimating the model parameters. In this work, we develop an estimation scheme for a very general form of this model and demonstrate its applicability to texture modeling applications. The relationship of the generalized LC models to other classes of RF models, namely the simultaneous autoregressive (SAR) and Markov random field (MRF) models, is shown. While it is known that the SAR model is a special case of the LC model, we show that the MRF model is also encompassed by this model. Consequently, the LC model may be considered as a generalization of the SAR and MRF models.

We consider the problem of maximum likelihood estimation applied to multispectral random field image models, specifically the multispectral simultaneous autoregressive (MSAR) model. Although previous work has provided least squares methods for parameter estimation, the maximum likelihood method often produces better results. For images with an assumed Gaussian distribution we develop effective procedures for calculating these estimates. Through a series of experiments using known random field models and natural texture samples, the effectiveness of the maximum likelihood approach is demonstrated.

A key component of the daily operation and planning activities of an electric utility is short-term load forecasting, i.e., the prediction of hourly loads (demand) for the next hour to several days out. The accuracy of such forecasts has significant economic impact for the utility. This paper describes a load forecasting system known as ANNSTLF (artificial neural-network short-term load forecaster) which has received wide acceptance by the electric utility industry and presently is being used by 32 utilities across the USA and Canada. ANNSTLF can consider the effect of temperature and relative humidity on the load. Besides its load forecasting engine, ANNSTLF contains forecasters that can generate the hourly temperature and relative humidity forecasts needed by the system. ANNSTLF is based on a multiple ANN strategy that captures various trends in the data. Both the first and the second generation of the load forecasting engine are discussed and compared. The building block of the forecasters is a multilayer perceptron trained with the error backpropagation learning rule. An adaptive scheme is employed to adjust the ANN weights during online forecasting. The forecasting models are site independent and only the number of hidden layer nodes of ANN's need to be adjusted for a new database. The results of testing the system on data from ten different utilities are reported.

We develop multispectral random field image models for use in image processing applications. The simultaneous autoregressive and Markov random field (MRF) models have been widely used in modeling intensity images. In this work we extend these models to include the more general multispectral case where images are represented by multiple intensity planes. In particular, we focus on the obvious application to color texture modeling using the RGB color model. For each model type we present the model equations, develop methods for synthesizing images based on these models and procedures for estimating the model parameters. In addition, the conditions necessary to ensure model validity are identified. We also provide experimental results which, substantiate the validity of these results. Color images synthesized from these models are shown to have the statistical characteristics implied by the model equations and parameters estimated from these images are very close to the known values from which the images were generated. In further experiments color random field models were fitted to natural texture samples. Images synthesized from these models are observed to be visually similar to the original images.

This paper describes a neural network (NN) based system for recognition and pose estimation of an unoccluded three-dimensional (3-D) object from any single two-dimensional (2-D) perspective view. The approach is invariant to translation, orientation, and scale. First, the binary silhouette of the object is obtained and normalized for translation and scale. Then, the object is represented by a set of rotation invariant features derived from the complex orthogonal pseudo-Zernike moments of the image. The recognition scheme combines the decisions of a bank of multilayer perceptron NN classifiers operating in parallel on the same data. These classifiers have different topologies and internal parameters, but are trained on the same set of exemplar perspective views of the objects. Next, two pose parameters, elevation and aspect angles, are obtained by a novel two-stage NN system consisting of a quadrant classifier followed by NN angle estimators. Performance is tested on clean and noisy data bases of military ground vehicles. Comparative studies with three other classifiers (a single NN, the weighted nearest-neighbor classifier, and a binary decision tree) are carried out.

In a landmark based medical image matching scheme, a coordinate transformation maps a set of landmark locations in one image onto a corresponding set in a second image. A novel approach to this transformation, the elastic body spline, is introduced. The development of the spline is outlined. An example of using the spline to match 3D magnetic resonance images (MRIs) of the breast is described. The elastic body spline is compared to the thin plate spline and, for 3D breast MRIs, is found to have superior performance as assessed by several measures of the similarity of the matched images.>

Many computer vision techniques (e.g., object recognition) use curvature as an image feature. Arbitrary, discrete, 3D surfaces are often encountered, for instance, in medical images (MRI is used as a particular example), but relatively few methods for estimating the curvature of these 3D surfaces have been proposed. Three such methods (the surface triangulation, cross patch, and partial derivative techniques) are described. These methods are used to locate curvature extrema (corners) in 3D images. The suitability of the methods for this application is evaluated and compared.

A system for recognition of segmented handwritten Persian/Arabic numerals irrespective of size and translation is developed. The image is represented by invariant features obtained from a new shadow coding scheme designed for the considered shapes. Classification is performed by a modified version of a four-layer probabilistic neural network (PNN) called the edited PNN (EPNN). Due to an editing and condensation procedure on the training samples, the EPNN has better performance and the network size is smaller. The performance of the system is evaluated on a database consisting of 2600 digits written by 10 different people. The obtained recognition accuracy is 97.8 percent. The developed system can process approximately two digits per second on a Intel 486 based PC with a 66 MHz clock.

This paper presents a stereo matcher inspired by the earlier work of Marr and Poggio (1976). Two major extensions are introduced: the algorithm is extended to gray-level images, and the inhibitory/excitatory weights of the model are learned rather than set a priori according to "uniqueness" and "continuity" constraints. Gray level stereo pairs of real scenes with known disparity maps are used to train the model. The trained system is successfully tested on other gray level stereo pairs of real scenes as well as a set of random dot stereograms. Performance is compared to a recent stereo matching algorithm.>

Random field models have been successfully utilized in many applications requiring texture synthesis, classification, and segmentation. This class of models assumes each image pixel can be represented as a function of neighboring pixels and an additive noise sample. The effectiveness of these models is highly dependent on the choice of neighbor sets. Current approaches to selecting neighbor sets are based on ad-hoc methods. In the paper a systematic method which selects neighbor sets based on the correlation structure of texture images is presented and evaluated.>

A neural network (NN) approach to the problem of steropsis is presented. The correspondence problem (finding the correct matches between pixels of the epipolar lines of the stereo pair from among all the possible matches) is posed as a noniterative many-to-one mapping. Two multilayer feedforward NNs are utilized to learn and code this nonlinear and complex mapping using the backpropagation learning rule and a training set. The first NN is a conventional fully connected net while the second is a sparsely connected NN with a fixed number of hidden layer nodes. All the applicable constraints are learned and internally coded by the NNs enabling them to be more flexible and more accurate than previous methods. The approach is successfully tested on several random-dot stereograms. It is shown that the nets can generalize their learned mappings to cases outside their training sets and to noisy images. Advantages over the Marr-Poggio algorithm are discussed, and it is shown that the NNs performances are superior.

The paper describes a vision system for automatic inspection of the connecting part of the wire bond of an IC where the wire connects to the bond pad on the chip. It considers a popular type of such bonds known as 'ball bond'. Using two-dimensional images taken from the top of the IC wafer, the system determines several geometric measures which are important in determining the quality of the bond. These measures include the boundary, length of major and minor axes of the best fitting ellipse and the center. The process utilizes automatic thresholding, morphological operations and geometric moments of the image. Success of the method is demonstrated through experimental studies on actual bonds.>

A nonparametric neural network technique for prediction of future values of a signal based on its past history is presented. A multilayer feed-forward neural network is used. It develops an internal model of the signal through a training operation involving past history of the considered signal. Training is performed using the back-propagation algorithm. The trained net is then used to do the forecast. Training is continued during operation to improve performance. The net performance is tested on signals generated by autoregressive (AR) models of orders two to ten, and results are compared to optimal forecasts.>

The problem of rotation-, scale-, and translation-invariant recognition of images is discussed. A set of rotation-invariant features are introduced. They are the magnitudes of a set of orthogonal complex moments of the image known as Zernike moments. Scale and translation invariance are obtained by first normalizing the image with respect to these parameters using its regular geometrical moments. A systematic reconstruction-based method for deciding the highest-order Zernike moments required in a classification problem is developed. The quality of the reconstructed image is examined through its comparison to the original one. The orthogonality property of the Zernike moments, which simplifies the process of image reconstruction, make the suggest feature selection approach practical. Features of each order can also be weighted according to their contribution to the reconstruction process. The superiority of Zernike moment features over regular moments and moment invariants was experimentally verified.>

A neural-network (NN) approach for translation-, scale-, and rotation-invariant recognition of objects is presented. The network utilized is a multilayer perceptron (MLP) classifier with one hidden layer. Backpropagation learning is used for its training. The image is represented by rotation-invariant features which are the magnitudes of the Zernike moments of the image. To achieve translation and scale invariancy, the image is first normalized with respect to these two parameters using its geometrical moments. The performance of the NN classifier on a database consisting of binary images of all English characters is reported and compared to those of nearest-neighbor and minimum-mean-distance classifiers. The results show that: (1) the MLP outperforms the other two classifiers, especially when noise is present; (2) the nearest-neighbor classifier performs about the same as the NN for the noiseless case; and (3) the Zernike-moment-based features possess strong class separability power.>

An algorithm for unsupervised texture segmentation is developed that is based on detecting changes in textural characteristics of small local regions. Six features derived from two, two-dimensional, noncausal random field models are used to represent texture. These features contain information about gray-level-value variations in the eight principal directions. An algorithm for automatic selection of the size of the observation windows over which textural activity and change are measured has been developed. Effects of changes in individual features are considered simultaneously by constructing a one-dimensional measure of textural change from them. Edges in this measure correspond to the sought-after textural edges. Experiments results with images containing regions of natural texture show that the algorithm performs very well.>

A method for recognizing an object in a binary image regardless of its orientation is discussed. The technique is also insensitive to slight deviation in shape and structure from a reference. The rotation-invariant features are the magnitudes of the Zernike moments of the image. Unlike classical moments, the Zernike moments are a mapping of the image onto a set of orthogonal basis functions, which gives them many useful properties. A novel synthesis-based approach for selection of these features is presented. Using this procedure, the discrimination power of features is evaluated by examining dissimilarities among images synthesized from them for different patterns. The method, applied to recognition of all English characters, yielded 95% accuracy.>

This paper presents a new model-based approach for texture classification which is rotation invariant, i.e., the recognition accuracy is not affected if the orientation of the test texture is different from the orientation of the training samples. The method uses three statistical features, two of which are obtained from a new parametric model of the image called a ``circular symmetric autoregressive model.'' Two of the proposed features have physical interpretation in terms of the roughness and directionality of the texture. The results of several classification experiments on differently oriented samples of natural textures including both microtextures and macrotextures are presented.