Indoor positioning in critical infrastructure is an emerging domain of research that is set to reinvent the way we navigate within GPS-deprived areas. The prolific pairing of smartphones and ubiquitous WiFi Access Points (APs) is expected to fuel the rapid adoption of this technology. However, maintaining long-term support for this technology undergoes unprecedented environmental changes over time owing to such factors as the gradual removal, addition, and replacement of APs. These changes lead to catastrophic degradation in quality-of-service. To overcome this challenge, this paper presents a novel set-and-forget framework using Laplacian Eigenmap manifold learning to enable long-term continual learning support for fingerprinting-based indoor positioning. An in-depth analysis of the proposed framework across a publicly available indoor UJI database shows improved positioning accuracy from 6% to 95% even in the presence of significant network configuration changes over time, when compared with the state-of-the-art.

Occupancy grids encode for hot spots on a map that is represented by a two dimensional grid of disjoint cells. The problem is to recursively update the probability that each cell in the grid is occupied, based on a sequence of sensor measurements from a moving platform. In this paper, we provide a new Bayesian framework for generating these probabilities that does not assume statistical independence between the occupancy state of grid cells. This approach is made analytically tractable through the use of binary asymmetric channel models that capture the errors associated with observing the occupancy state of a grid cell. Binary-valued measurement vectors are the thresholded output of a sensor in a radar, sonar, or other sensory system. We compare the performance of the proposed framework to that of the classical formulation for occupancy grids. The results show that the proposed framework identifies occupancy grids with lower false alarm and miss detection rates, and requires fewer observations of the surrounding area, to generate an accurate estimate of occupancy probabilities when compared to conventional formulations.

This paper addresses the problem of performance prediction and estimation for a multichannel coherence detector. The saddlepoint approximation is employed to approximate the empirical null distribution of the test statistics for the purpose of predicting the false alarm performance. The ability of the proposed method to achieve a prescribed false alarm rate is then examined for different background conditions using a Synthetic Aperture Sonar (SAS) imagery dataset.

Dictionary learning methods, such as K-SVD, have recently been introduced to learn a specific dictionary matrix that best fits a set of training data vectors. These methods are flexible in that any preferred pursuit method of sparse coding can be used to represent the data. Incremental in-situ learning using unseen data, however, requires complete retraining and hence increased computational efforts. In this paper, we present a new recursive algorithm to jointly perform dictionary learning and sparse coding using a fast orthogonal matching pursuit (OMP) when unseen data samples are encountered. The incremental in-situ learning is accomplished while maintaining the performance and with substantially reduced computations. Simulation results are presented to show the effectiveness and benefits of the proposed algorithms.

This paper introduces a kernel machine for multiclass discrimination where the scoring function for each class is constructed using a linear combination over a predefined diverse library of kernel functions. The scoring function is built using an expanded set of the kernel library hence increasing the number of degrees of freedom to analyze the information content of each data sample. To choose the smallest set of kernels that best match desirable first-order moment properties of the class-conditional distribution a regularized linear least-squares problem is solved. The proposed multi-kernel machine is then demonstrated and benchmarked against similar techniques which rely on the use of a single kernel using a satellite imagery dataset for the purposes of discriminating among several vegetation and soil types.

This letter considers the problem of threshold selection for a correlation test among multiple (≥2) random vectors. The generalized likelihood ratio test (GLRT) for this problem uses a generalized Hadamard ratio to test for block diagonality in a composite covariance matrix. As the number of realizations used to estimate the composite covariance matrix grows large, the null distribution of the likelihood ratio statistic converges to a chi-squared distribution which can be used to prescribe thresholds needed to achieve a desired false alarm rate in high sample support situations. However, this asymptotic distribution can be slow to converge, making its use dubious in many practical scenarios. To address this problem, this letter uses saddlepoint approximations for the null distribution of the generalized Hadamard ratio. Simulations are provided to demonstrate the saddlepoint approximation's ability to achieve a desired false alarm probability, even in situations with low sample support.

This paper proposes a supervised approach for analysis of high-dimensional data using low-dimensional submanifolds. This method offers many useful properties. Using first order approximation for the given nonlinear mapping, we introduce a locally linear model. This model is such that it minimizes the local approximation error resulted by mapping to a local subspace during the learning. Additionally, the proposed method preserves local data energy to conserve local topology. Finally, this method guarantees the separability of the mapped data for different data classes. Two different approaches used for this aim, Linear Discriminant Analysis (LDA) and Regularized Maximum Margin Criterion (RMMC). Having those local feature-domain data, the whole feature domain data can be estimated in MMSE sense. The performance of this method is demonstrated on a sonar imagery dataset for classification of underwater objects.

This paper considers the problem of testing for the independence among multiple random vectors with each random vector representing a time series captured at one sensor. Implementing the Generalized Likelihood Ratio Test involves testing the null hypothesis that the composite covariance matrix of the channels is block-diagonal through the use of a generalized Hadamard ratio. These results are then extended to the problem of detecting the presence of correlated time series when several observers each employ an array of sensors. Assuming wide-sense stationary processes in both time and space, results on large block-Toeplitz matrices suggest the use of a broadband integral of a frequency-wavenumber dependent Hadamard ratio as an alternative test statistic.

K-SVD method has recently been introduced to learn a specific dictionary matrix that best fits a set of training data vectors. K-SVD is flexible in that any preferred pursuit method of sparse coding can be used to represent the data. In this paper, we show how K-SVD method can be used in conjunction with a fast orthogonal matching pursuit implemented using orthogonal projection updating. Geometric interpretation of this learning is also presented. The method was then applied to underwater target detection problem using a dual-channel sonar imagery data.

This paper considers sequential detection and classification of multiple transient signals from vector observations corrupted with additive noise and multiple types of structured interference. Sparse approximations of observations are found to facilitate computation of the likelihood of each signal model without relying on restrictive assumptions concerning the distribution of observations. Robustness to interference may be incorporated by virtue of the inherent separation capabilities of sparse coding. Each signal model is characterized by a Bayesian Network, which captures the temporal dependency structure among coefficients in successive sparse approximations under the associated hypothesis. Generalized likelihood ratios tests may then be used to perform signal detection and classification during quiescent periods, and quiescent detection whenever a signal is present. The results of applying the proposed method to a national park soundscape analysis problem demonstrate its practical utility for detecting and classifying real acoustical sources present in complex sonic environments.

Acoustic travel-time tomography of the atmosphere is a nonlinear inverse problem which attempts to reconstruct temperature and wind velocity fields in the atmospheric surface layer using the dependence of sound speed on temperature and wind velocity fields along the propagation path. This paper presents a new statistical-based acoustic travel-time tomography algorithm based on unscented Kalman filter (UKF) which is capable of reconstructing and tracking temperature and wind velocity fields (state variables) within a specified investigation area. The method exploits an iterative ray-tracing algorithm to handle situations when straight-ray assumption no longer holds. The observations used in the UKF process consists of the acoustic travel times computed for every pair of transmitter/reciever nodes deployed in the investigation area. A first-order spatial-temporal autoregressive model is used to account for state evolution in the UKF. To evaluate the performance of the UKF-based acoustic tomography method, 2-D fractal Brownian motion is used to generate synthetic temperature and wind velocity fields with spatial and temporal resolution of 1 m and 12 s, respectively. The UKF-based acoustic tomography algorithm is then compared to the well-known time-dependent stochastic inversion method. The results reveal the effectiveness of the proposed method for accurate and fast reconstruction of temperature and wind velocity fields.

This paper considers the problem of testing for the independence among multiple (≥ 2) random vectors with each random vector representing a time series captured at one sensor. Implementing the Generalized Likelihood Ratio Test involves testing the null hypothesis that the composite covariance matrix of the channels is block-diagonal through the use of a generalized Hadamard ratio. Using the theory of linear prediction and its connection with Gram determinants, it is shown that this generalized Hadamard ratio can be written as a product of scalars which are independently drawn from a beta distribution under the null hypothesis. This result is useful from a Monte Carlo analysis standpoint in that it is much more computationally efficient to form a product of scalar beta random variables than it is to compute the determinant of complex Wishart matrices.

This paper introduces a new framework for detecting, classifying, and estimating the signatures of multiple transient acoustical sources from a time-transform representation (TTR) of an audio waveform. A TTR is a vector observation sequence containing the coefficients of consecutive windows of data with respect to known sampled basis waveforms. A set of likelihood ratio tests is hierarchically applied to each time slice of a TTR to detect and classify signals in the presence of interference. Since the signatures of each acoustical event typically span several adjacent dependent observations, a Kalman filter is used to generate the parameters necessary for computing the likelihood values. The experimental results of applying the proposed method to a problem of detecting and classifying man-made and natural transient acoustical events in national park soundscape recordings attest to its effectiveness at performing the aforementioned tasks.

This paper introduces a method for in-situ underwater target classification, based on an image retrieval system, that can be implemented using a simple two-layer kernel-based network. This system incorporates a learning mechanism that captures new information for discriminating between objects in different classes or within the same class from a set of input-output pairs with associated confidence scores. A strategy to select the most informative patterns for optimal parameter adaptation during in-situ learning is also described. The system is then tested on a database of synthetically generated sonar images. The ability of the system to correctly classify images containing objects in different environmental and operating conditions than those used for original training, as well as its ability to incorporate new object types without perturbing the classification performance on other object types are demonstrated.

New fusion methods for an operationally adaptive (OA) system for prediction of acoustic transmission loss (TL) in the atmosphere are developed in this paper. The OA system uses expert neural network predictors, each corresponding to a specific range of source elevation. The outputs of the expert predictors are combined using two new nonlinear fusion methods. Using this prediction methodology the computational intractability of traditional acoustic propagation models is eliminated. The proposed fusion methods are tested on a synthetically generated acoustic data set for a wide range of geometric, source, and environmental conditions.

This paper presents a new method for multiaspect/ping classification of underwater objects using sonar data. This system uses decision feedback to form a likelihood ratio for making a high confidence final decision based not only upon the data of the current ping but also the final decisions made at several previous pings. The system is then applied to an underwater target classification problem. Test results on a buried object scanning sonar (BOSS) database collected for different objects and in different conditions show the promise of the proposed method for multi-aspect underwater target discrimination.

This paper introduces a new target detection method for multiple disparate sonar platforms. The detection method is based upon multi-channel coherence analysis (MCA) framework which allows one to optimally decompose the multichannel data to analyze their linear dependence or coherence. This decomposition then allows one to extract MCA features which can be used to discriminate between two hypotheses, one corresponding to the presence of a target and one without, through the use of the log-likelihood ratio. Test results of the proposed detection system were applied to a data set of underwater side-scan sonar imagery provided by the Naval Surface Warfare Center (NSWC), Panama City. This database contains data from 4 disparate sonar systems, namely one high frequency (HF) sonar and three broadband (BB) sonars coregistered over the same area on the sea floor. Test results illustrate the effectiveness of the proposed multi-platform detection system in terms of probability of detection, false alarm rate, and receiver operating characteristic (ROC) curves.

This paper uses the canonical correlation decomposition (CCD) framework to investigate the spatial correlation of sources captured using two spatially separated sensor arrays. The relationship between the canonical correlations of the observed signals and the spatial correlation coefficients of the source signals are first derived, including an analysis of the changes seen in this relationship under certain noise level and array geometry assumptions. Additionally, simulation results are presented that demonstrate the effects of different noise levels and array geometries on the canonical correlations for the case of two uniform linear sparse arrays.

This paper presents an adaptable content-based image retrieval (CBIR) system developed using regularization theory, kernel-based machines, and Fisher information measure. The system consists of a retrieval subsystem that carries out similarity matching using image-dependant information, multiple mapping subsystems that adaptively modify the similarity measures, and a relevance feedback mechanism that incorporates user information. The adaptation process drives the retrieval error to zero in order to exactly meet either an existing multiclass classification model or the user high-level concepts using reference-model or relevance feedback learning, respectively. To facilitate the selection of the most informative query images during relevance feedback learning a new method based upon the Fisher information is introduced. Model-reference and relevance feedback learning mechanisms are thoroughly tested on a domain-specific image database that encompasses a wide range of underwater objects captured using an electro-optical sensor. Benchmarking results with two other relevance feedback learning methods are also provided.

The problem of source localization using distributed acoustic sensor networks is considered in this paper. The acoustic sensors detect and estimate the time of arrival (TOA) of the acoustic transient emitted from a source of interest. The estimated TOA estimates are then transmitted to a base station for collaborative source localization. In realistic scenarios, a subset of the TOA's could typically be erroneous owing to various detrimental factors. The standard least squares-based methods totally fail in presence of even only one outlier or erroneous TOA. By modeling the TOA estimation error as Cauchy-Lorentz distribution, a robust source localization algorithm is derived. Numerical experiments are provided to show the robustness and accuracy of localization of the proposed algorithm.

In this paper, a new collaborative multi-aspect classification system (CMAC) is introduced. CMAC utilizes a group of collaborative decision-making agents capable of producing a high-confidence final decision based on features obtained over multiple aspects. This system is then applied to a buried underwater target classification problem. The results show that CMAC provides excellent multi-ping classification of mine-like objects while simultaneously reducing the number of false alarms compared to a multi-ping decision-level fusion classifier.

An operationally adaptive (OA) system for prediction of acoustic transmission loss (TL) in the atmosphere is developed in this paper. This system uses expert neural network predictors, each corresponding to a specific range of source elevation. The outputs of the expert predictors are combined using a performance-aware weighting mechanism and a nonlinear fusion system. Using this prediction methodology the computational intractability of traditional acoustic models is eliminated. The proposed system is tested on a synthetically generated acoustic data set for a wide range of geometric, source, and environmental conditions.

An iterative learning algorithm for multi-channel coherence analysis (MCA) is developed in this paper. MCA is an extension of the well known canonical correlation analysis (CCA) that allows for more than two data channels to be analyzed. The many applications of CCA have motivated this extension to exploit the linear relationship between many data channels. This paper discusses fundamental differences between the two analysis techniques while reviewing the standard method for performing MCA. Discussion on why MCA correlations are not deemed "canonical" as they are in the two-channel case of CCA is also provided. The developed iterative learning for MCA is then demonstrated and its performance evaluated on a synthesized data set.

State-of-the-art cloud products are typically generated using scientific polar orbiting satellites such as the Moderate Resolution Imaging Spectroradiometer (MODIS). However, they do not allow for observation of the same region at a regular temporal frequency, rendering them ineffectual for nowcasting problems. Operational satellites such as Meteosat-8 SEVIRI, in contrast, are geostationary and provide continual data at a regular temporal frequency over a much larger region. MODIS-like cloud products cannot be directly generated from operational satellites as they typically have a smaller number of spectral bands and different wavelengths and spatial resolution. This paper applies the canonical coordinate decomposition method to estimate scientific cloud products using imagery from operational satellites. Using the proposed method features of the Meteosat-8 imagery data that are maximally coherent with the data from the MODIS are generated. These features are temporally updated at times and locations where MODIS data are unavailable using the alternating block power method. A subset of the canonical coordinates of Meteosat-8 SEVIRI is then used to create MODIS-like cloud products using several neural networks. The quality of the generated cloud products and their temporal consistency have been demonstrated on several data sets from July 2004. A benchmarking with an independent Meteosat-8-based algorithm is also provided, which shows the promise of our approach in generating MODIS-like cloud products

This paper introduces a new connectionist network for certain domain-specific text-retrieval and search applications with expert end users. A new model reference adaptive system is proposed that involves three learning phases. Initial model-reference learning is first performed based upon an ensemble set of input-output of an initial reference model. Model-reference following is needed in dynamic environments where documents are added, deleted, or updated. Relevance feedback learning from multiple expert users then optimally maps the original query using either a score-based or a click-through selection process. The learning can be implemented, in regression or classification modes, using a three-layer network. The first layer is an adaptable layer that performs mapping from query domain to document space. The second and third layers perform document-to-term mapping, search/retrieval, and scoring tasks. The learning algorithms are thoroughly tested on a domain-specific text database that encompasses a wide range of Hewlett Packard (HP) products and for a large number of most commonly used single- and multiterm queries.

In this paper, we introduce a non-iterative nonlinear kernel Wiener filtering method using kernel canonical correlation analysis (CCA) framework. This approach is based upon the theory of reproducing kernel Hubert spaces. A method is proposed to find approximate Wiener filtered signal in the original signal space by solving an optimization problem in the higher dimensional space. Unlike the conventional iterative approaches which rely on nonlinear optimization problem, our proposed method directly finds the pre-image using distance constraints in the higher mapped domain. The signal estimation and reconstruction capability of the new method is demonstrated and benchmarked on the United States Postal Service (USPS) digits database. Moreover, a comparison with the conventional kernel Wiener filter is presented

Multichannel Canonical Correlation Analysis (MCCA) is used in this paper for feature extraction from multiple sonar returns off of buried underwater objects using data collected by the new generation Buried Object Scanning Sonar (BOSS) system. Comparisons are made between the classification results of features extracted by the proposed algorithm and those extracted by the two-channel Canonical Correlation Analysis (CCA) algorithm. This study compares different feature extraction and classification algorithms, and the results are presented in terms of confusion matrices. The results show that MCCA yields higher correct classification rates than CCA while reducing the classifier's structural complexity.

An environmentally adaptive system for prediction of acoustic transmission loss (TL) in the atmosphere is developed in this paper. This system uses expert neural network predictors, each corresponding to a specific environmental condition. The outputs of the expert predictors are combined using a fuzzy confidence measure and a non-linear fusion system. Using this prediction methodology the computational intractability of traditional acoustic models is eliminated. The proposed system is tested on a synthetic acoustic data set for a wide range of geometric, source, and environmental conditions.

A new feature adaptation mechanism is proposed in this paper that captures the relevance feedback information provided by the expert users. This relevance information is retained for future usage and subsequently made available to other users. The search and retrieval processes are implemented through a two-layer connectionist network structure and the relevance feedback learning is incorporated by appropriately modifying the network structure. The developed algorithm is tested on an electro-optical imagery database collected from different underwater mine-like and non-mine-like objects.

In this paper, relation between multi-class linear and kernel Fisher discriminant analysis (FDA) and linear and kernel canonical correlation analysis (CCA) is established. It is shown that in a multi-class classification problem, the CCA between feature vectors (or a nonlinearly mapped version of them) as one-channel and the class label vectors as the second channel is equivalent to multi-class FDA. The multi-class Fisher distance is found to be decomposed into a sum of terms, each of which is determined by a canonical correlation. This result is extended to the kernel formulation without explicit computation of the nonlinear mappings. A simple example is presented to numerically verify the results.

This paper presents a new multi-aspect pattern classification method using hidden Markov models (HMMs). Models are defined for each class, with the probability found by each model determining class membership. Each HMM model is enhanced by the use of a multilayer perception (MLP) network to generate emission probabilities. This hybrid system uses the MLP to find the probability of a state for an unknown pattern and the HMM to model the process underlying the state transitions. A new batch gradient descent-based method is introduced for optimal estimation of the transition and emission probabilities. A prediction method in conjunction with HMM model is also presented that attempts to improve the computation of transition probabilities by using the previous states to predict the next state. This method exploits the correlation information between consecutive aspects. These algorithms are then implemented and benchmarked on a multi-aspect underwater target classification problem using a realistic sonar data set collected in different bottom conditions.

This paper introduces a new connectionist network for large-scale text retrieval applications. A learning mechanism is proposed to optimally map the original query using relevance feedback from multiple expert users. The query mapping not only meets the requirements of the expert users but also preserves the positions and ranks of other relevant documents. An updating algorithm is also proposed to incorporate new documents (or delete the obsolete ones) into the system either one-by-one or in a batch mode without requiring to retrain the system. The algorithms are successfully tested on a large database and for a large number of most commonly used single-term or multi-terms queries.

A kernel-based formulation for decomposing nonlinear maps of two data channels into their canonical coordinates is derived. Each data channel is implicitly mapped to a high dimensional feature space defined by a nonlinear kernel. The canonical coordinates of the nonlinear maps are then found by transforming the kernel maps with the eigenvector matrices of a coupled asymmetric generalized eigenvalue problem. This generalized eigenvalue problem is constructed in the explicit space of kernel maps. The measures of linear dependence and coherence between the nonlinear maps of the channels are also presented. These measures may be determined in the kernel domain, without explicit computation of the nonlinear mappings. A numerical example is also presented.

Classification of underwater targets from the acoustic backscattered signals is considered here. Several different classification algorithms are tested and benchmarked not only for their performance but also to gain insight to the properties of the feature space. Results on a wideband 80-kHz acoustic backscattered data set collected for six different objects are presented in terms of the receiver operating characteristic (ROC) and robustness of the classifiers wrt reverberation.

This paper presents a new temporally adaptive classification system for multispectral images. A spatial-temporal adaptation mechanism is devised to account for the changes in the feature space as a result of environmental variations. Classification based upon spatial features is performed using Bayesian framework or probabilistic neural networks (PNNs) while the temporal updating takes place using a spatial-temporal predictor. A simple iterative updating mechanism is also introduced for adjusting the parameters of these systems. The proposed methodology is used to develop a pixel-based cloud classification system. Experimental results on cloud classification from satellite imagery are provided to show the usefulness of this system.

A network structure for canonical coordinate decomposition is presented. The network consists of two single-layer linear subnetworks that together extract the canonical coordinates of two data channels. The connection weights of the networks are trained by a stochastic gradient descent learning algorithm. Each subnetwork features a hierarchical set of lateral connections among its outputs. The lateral connections perform a deflation process that subtracts the contribution of the already extracted coordinates from the input data subspace. This structure allows for adding new nodes for extracting additional canonical coordinates without the need for retraining the previous nodes. The performance of the network is evaluated on a synthesized data set.

Hybrid hidden Markov models (HMM) and multi-layer (MLP) neural networks have been applied great success in speech recognition problems. The hybrid system can be applies to sequence classification problems, where multiple looks at an object are used to determine class membership. This presents a utility to perform feature-level fusion in such problems. A new gradient descent algorithm is employed to find optimal parameters within the HMM/MLP model. This scheme has been applied to a data set which contains sonar backscattered signals for four underwater objects for classification as mine-like or non-mine-like.

A two-channel temporal updating system is presented, which accounts for feature changes in the visible and infrared satellite images. The system uses two probabilistic neural network classifiers and a context-based predictor to perform continuous cloud classification during the day and night. Test results for 27 h of continuous classification and updating are presented on a sequence of Geostationary Operational Environmental Satellite 8 images. Further test results of the system on two new sets of data with 1-2 weeks time difference are also presented that show the potential of this system as an operational continuous cloud classification system.

A new adaptive underwater target classification system to cope with environmental changes in acoustic backscattered data from targets and nontargets is introduced. The core of the system is the adaptive feature mapping that minimizes the classification error rate of the classifier. The goal is to map the feature vector in such a way that the mapped version remains invariant to the environmental changes. A K-nearest neighbor (K-NN) system is used as a memory to provide the closest matches of an unknown pattern in the feature space. The classification decision is done by a backpropagation neural network (BPNN). Two different cost functions for adaptation are defined. These two cost functions are then combined together to improve the classification performance. The test results on a 40-kHz linear FM acoustic backscattered data set collected from six different objects are presented. These results demonstrate the effectiveness of the adaptive system versus nonadaptive system when the signal-to-reverberation ratio (SRR) in the environment is varying.

Stereo image compression involves estimating the disparity vectors that represent the amount of binocular parallax. The mismatching problems between the left and right images greatly impact the accuracy of the reconstructed image, and hence the visual effects of the reproduced 3-D image. This paper presents a new method for compensating the mismatching effects in stereo image pairs. This 2-D filtering-based scheme uses a sequential orthogonal subspace updating (SOSU) process to project an image block onto a subset of best-basis vectors. The basis vectors are selected one by one from the neighboring blocks, as well as some typical edge blocks, forming an image-dependent set of basis vectors. This leads to the optimal representation of an image block with fewer coefficients. Simulation results on two different image pairs demonstrate the effectiveness of the SOSU scheme when compared to those of the standard least squares 2-D filtering and the hybrid disparity-compensated discrete cosine transform residual encoding schemes.

A novel temporal updating approach for probabilistic neural network classifiers was developed by Tian et al. (2000) to account for temporal changes of spectral and temperature features of clouds in the visible and infrared GOES 8 (Geostationary Operational Environmental Satellite) imagery data. In this paper, a new method referred to as moving singular value decomposition (MSVD) is introduced to improve the classification rate of the boundary blocks or blocks containing cloud types with non-uniform texture. The MSVD method is then incorporated into the temporal updating scheme and its effectiveness is demonstrated on several sequences of GOES 8 cloud imagery data. These results indicate that the incorporation of the new MSVD improves the overall performance of the temporal updating process by almost 10%

Presents a training algorithm for probabilistic neural networks (PNN) using the minimum classification error (MCE) criterion. A comparison is made between the MCE training scheme and the widely used maximum likelihood (ML) learning on a cloud classification problem using satellite imagery data.

A two-dimensional (2-D) least squares (LS)-based filtering scheme for high fidelity stereo image compression applications is introduced in this correspondence. This method removes the effects of mismatching in a stereo image pair by applying the left image as the reference input to a 2-D transversal filter while the right image is used as the desired output. The weights of the filter are computed using a block-based LS method. A reduced order filtering scheme is also introduced to find the optimum number of filter coefficients. The principal coefficients and the disparity vectors are used together with left image to reconstruct the right image at the receiver, The proposed schemes are examined on a real stereo image pair for 3D-TV applications and the results were benchmarked against those of the block-matching method.

In this paper, a new subband-based classification scheme is developed for classifying underwater mines and mine-like targets from the acoustic backscattered signals. The system consists of a feature extractor using wavelet packets in conjunction with linear predictive coding (LPC), a feature selection scheme, and a backpropagation neural-network classifier. The data set used for this study consists of the backscattered signals from six different objects: two mine-like targets and four nontargets for several aspect angles. Simulation results on ten different noisy realizations and for signal-to-noise ratio (SNR) of 12 dB are presented. The receiver operating characteristic (ROC) curve of the classifier generated based on these results demonstrated excellent classification performance of the system. The generalization ability of the trained network was demonstrated by computing the error and classification rate statistics on a large data set. A multiaspect fusion scheme was also adopted in order to further improve the classification performance.

In cloud classification from satellite imagery, temporal change in the images is one of the main factors that causes degradation in the classifier performance. In this paper, a novel temporal updating approach is developed for probabilistic neural network (PNN) classifiers that can be used to track temporal changes in a sequence of images. This is done by utilizing the temporal contextual information and adjusting the PNN to adapt to such changes. Whenever a new set of images arrives, an initial classification is first performed using the PNN updated up to the last frame while at the same time, a prediction using Markov chain models is also made based on the classification results of the previous frame. The results of both the old PNN and the predictor are then compared. Depending on the outcome, either a supervised or an unsupervised updating scheme is used to update the PNN classifier. Maximum likelihood (ML) criterion is adopted in both the training and updating schemes. The proposed scheme is examined on both a simulated data set and the Geostationary Operational Environmental Satellite (GOES) 8 satellite cloud imagery data. These results indicate the improvements in the classification accuracy when the proposed scheme is used.

This paper presents a new training algorithm for probabilistic neural networks (PNNs) using the minimum classification error (MCE) criterion. A comparison is made between the MCE training scheme and the widely used maximum likelihood learning on a cloud classification problem using satellite imagery data.

In this paper, a new feature mapping scheme is presented to cope with environmental and target signature changes for underwater target classification. A wavelet packet-based feature extraction scheme is used in conjunction with the linear prediction coding (LPC) scheme as the front-end processor. The core of the system is the adaptive feature mapping subsystem that minimizes the classification error of the classifier. The extracted feature vector is mapped by the resultant transformation matrix in such a way that the mapped version remains invariant to the environmental and sensory changes. The feedback to the adaptation mechanism is provided by a k-nearest neighbor classifier. The test results on 40 kHz linear FM acoustic backscattered data collected for six different objects are presented The effectiveness of the adaptive system vs. nonadaptive one is demonstrated for several signal-to-noise ratio (SNR) conditions.

In this paper, a new neural network-based sequential Bayes classifier is developed for target classification exploring the target's dynamic behavior. The system consists of a feature extractor, a neural network directed conditional probability generator and a sequential Bayes classifier. The velocity and curvature sequences extracted from each trade are used as the primary features. Several hidden states are used to train the neural network, the output of which is the conditional probability of occurring the hidden states given the observations. These conditional probabilities are then used as the inputs to the Bayes classifier to make the classification. The classification results are updated regressively whenever a new scan of data is received. Simulation results on both clean tracks and heavily cluttered infrared (IR) satellite images are presented to demonstrate the effectiveness of the proposed methods.

An automated approach for template-free identification of partially occluded objects is presented. The contour of each relevant object in the analyzed scene is modeled with an approximating polygon whose edges are then projected into the Hough space. A structurally adaptive self-organizing map neural network generates clusters of collinear and/or parallel edges, which are used as the basis for identifying the partially occluded objects within each polygonal approximation. Results on a number of cases under different conditions are provided.

A neural network-based scheme for decision directed edge-adaptive Kalman filtering is introduced in this work. A backpropagation neural network makes the decisions about the orientation of the edges based on the information in a window centered at the current pixel being processed. Then based upon the neural network output an appropriate image model which closely matches the local statistics of the image is chosen for the Kalman filter. This prevents the oversmoothing of the edges, which would have otherwise been caused by the standard Kalman filter. Simulation results are presented which show the effectiveness of the proposed scheme.

The problem of cloud data classification from satellite imagery using neural networks is considered in this paper. Several image transformations such as singular value decomposition (SVD) and wavelet packet (WP) were used to extract the salient spectral and textural features attributed to satellite cloud data in both visible and infrared (IR) channels. In addition, the well-known gray-level cooccurrence matrix (GLCM) method and spectral features were examined for the sake of comparison. Two different neural-network paradigms namely probability neural network (PNN) and unsupervised Kohonen self-organized feature map (SOM) were examined and their performance were also benchmarked on the geostationary operational environmental satellite (GOES) 8 data. Additionally, a postprocessing scheme was developed which utilizes the contextual information in the satellite images to improve the final classification accuracy. Overall, the performance of the PNN when used in conjunction with these feature extraction and postprocessing schemes showed the potential of this neural-network-based cloud classification system.

Prospective authors are requested to submit new, unpublished manuscripts for inclusion in the upcoming event described in this call for papers.

In processing of sonar data, beamforming process plays a central role in reducing the effects of the surface and bottom reverberation. In shallow water environments where the reverberation is dominant, target detection from the beamformed results is not effective and may lead to significantly high false alarm rate. This paper presents a novel approach for post processing sonar beamformed imagery in order to improve the detectability of the targets while substantially reducing the occurrence of the false detection. This is done using the recursive high order correlation (RHOC) method which exploits the spatial-temporal correlation between consecutive pings of the beamformed images. Test results on several sets of sonar data show the great efficiency and power of the proposed method especially in very high cluttered environments.

This paper presents a 2-D filtering scheme for stereo image compression using orthogonal subspace projection. To provide more candidate blocks for input data, the support region for input data is extended in the reference image. In addition, edge blocks are added to the candidate input blocks in order to provide better compensation ability for edges and boundaries of objects. The best blocks for input data are selected one by one in order of importance to reconstruct the desired block using the Gram-Schmidt orthogonalization algorithm. Simulation results exhibit excellent performance of the proposed scheme when compared to those of the standard block-matching and least-squares(LS)-based 2-D filtering schemes.

This paper presents a 2-D filtering scheme for stereo image compression using sequential orthogonal subspace updating (SOSU). The basis vectors for this representation include an extended set of blocks in the support region as well as some edge blocks for providing better compensation ability for edges and boundaries of objects. The desired image block is then projected onto a selected subset of these vectors for optimal expansion. The basis vectors are selected one by one in order of importance using the Gram-Schmidt (GS) orthogonalization procedure. The proposed scheme is inspired from the least-squares (LS) based 2-D filtering scheme of Seo, Azimi-Sadjadi and Tian (see Intl. Conf. Image Processing, vol.3, p.260-3, 1997) while circumventing its limitations by enhancing the compensation ability for the mismatching problems in stereo image pairs. Simulation results demonstrate excellent performance of the proposed scheme when compared to those of the hybrid disparity-compensated discrete cosine transform (DCT) residual coding and the LS-based 2-D filtering schemes.

This paper introduces a new system for real-time detection and classification of arbitrarily scattered surface-laid mines from multispectral imagery data of a minefield. The system consists of six channels which use various neural-network structures for feature extraction, detection, and classification of targets in six different optical bands ranging from near UV to near IR. A single-layer autoassociative network trained using the recursive least square (RLS) learning rule was employed in each channel to perform feature extraction. Based upon the extracted features, two different neural-network architectures were used and their performance was compared against the standard maximum likelihood (ML) classification scheme. The outputs of the detector/classifier network in all the channels were fused together in a final decision-making system. Two different final decision making schemes using the majority voting and weighted combination based on consensual theory were considered. Simulations were performed on real data for six bands and on several images in order to account for the variations in size, shape, and contrast of the targets and also the signal-to-clutter ratio. The overall results showed the promise of the proposed system for detection and classification of mines and minelike tagets.

A block-matching method is generally used for the disparity estimation as well as motion estimation applications. However, this method suffers from limitations such as blocking artifacts on the reconstructed images and a lack of compensation ability for the mismatching areas. A new disparity estimation scheme using 2-D filtering is proposed to provide more accurate estimates of the disparity vector and better compensation ability. This is accomplished by applying the left image to the reference input of the filter while using the right image as the desired output. The goal is to provide the best matching for the right image using the filter output instead of the left image directly as used in the block-matching method. A reduced order filtering scheme is also introduced to minimize the number of filter coefficients for the reconstruction. The reconstructed images generated based upon the estimated disparity vectors and some principal filter coefficients exhibited excellent quality.

Automatic cloud classification of satellite imagery can be of great help to meteorological studies. A neural network-based cloud classification system is developed and introduced. Several image transformation schemes such as wavelet transform (WT) and singular value decomposition (SVD) are used to extract the salient textural feature of the data and is then compared with those of the well-known gray-level co-occurrence matrix (GLCM) approach. Two different neural network paradigms namely the probability neural network (PNN) and the unsupervised Kohonen (1990) self-organized feature map (SOM) are chosen and examined. The performance of the proposed cloud classification system is benchmarked on the Geostationary Operational Environmental Satellite (GOES) 8 data set and promising results have been achieved.

Describes an automated approach for analyzing complex scenes obtained from digitized scanning electron microscopy (SEM) images of lung bio-dissolution preparations. Such images are used in the evaluation of airborne fiberglass distributions. They present varying densities of isolated, overlapping, and crossing fibers and other objects that should be individually identified and measured. A polygonal approximation model for each detected object in the scene is used to obtain a simplified description. After detecting complex objects, such as crossing fibers, an unsupervised neural network is used to resolve the ambiguities and separate the objects. The method performs satisfactorily under different conditions of fiber density and background clutter.

An adaptive learning approach for the computation of the coefficients of the generalized nonorthogonal 2-D Gabor (1946) transform representation is introduced. The algorithm uses a recursive least squares (RLS) type algorithm. The aim is to achieve minimum mean squared error for the reconstructed image from the set of the Gabor coefficients. The proposed RLS learning offers better accuracy and faster convergence behavior when compared with the least mean squares (LMS)-based algorithms. Applications of this scheme in image data reduction are also demonstrated.

The problem of underwater target detection and classification from acoustic backscatter is the central focus of this paper. A new approach for identifying the presence of resonance in the acoustic backscatter from unknown elastic targets by isolating the resonance part from the specular contribution is developed. The method allows for characterization of both the time history and time-scale of the resonant contribution of the target directly from the acoustic backscatter. An adaptive transversal filter structure is used to estimate the specular part of the backscatter and consequently the error signal would provide an estimate of the resonance part. This scheme does not require any underlying model assumption about the elastic return and further can be applied to targets of unknown geometry and thickness. The adaptation rule is based upon fast recursive least squares (RLS) learning. Test results on acoustic data are presented which indicate the effectiveness of the proposed approach.

A new neural network-based approach is introduced for recursive computation of the principal components of a stationary vector stochastic process. The neurons of a single-layer network are sequentially trained using a recursive least squares squares (RLS) type algorithm to extract the principal components of the input process. The optimality criterion is based on retaining the maximum information contained in the input sequence so as to be able to reconstruct the network inputs from the corresponding outputs with minimum mean squared error. The proof of the convergence of the weight vectors to the principal eigenvectors is also established. A simulation example is given to show the accuracy and speed advantages of this algorithm in comparison with the existing methods. Finally, the application of this learning algorithm to image data reduction and filtering of images degraded by additive and/or multiplicative noise is considered.

A simple neural network-based approach is introduced in this paper which allows the computation of the coefficients of the generalized non-orthogonal 2-D Gabor transform representation. The network is trained using a recursive least squares (RLS) type algorithm. This RLS learning algorithm offers better accuracy and faster convergence when compared to the least mean squares (LMS) based algorithms. The aim is to achieve minimum mean squared error for the reconstructed image from the set of the Gabor coefficients. Application of this scheme in image data reduction is demonstrated in the simulation results.>

Parallel and pipeline structures for real-time 2D recursive filtering are presented. For general scalar 2D recursive filters, a 2D multiple-interleaved pipeline architecture is introduced that is compatible with the nature of the image-scanning scheme. Using this structure, the sampling period can be a fraction of the time needed for one scalar addition operation and the delay is only a few samples. In addition, this structure does not require any I/O buffers for real-time implementation.>

The development of a neural-network-based classifier for classifying three distinct scenes (urban, park, and water) from several polarized SAR images of the San Francisco Bay area is discussed. The principal components (PC) scheme or Karhunen-Loeve transform is used to extract the salient features of the input data, and to reduce the dimensionality of the feature space prior to the application to the neural networks. Using the PC scheme along with the polarized images used in the present study led to substantial improvements in the classification rates when compared with previous studies. When a combined polarization architecture was used, the classification rate for water, urban, and park areas improved to 100%, 98.7%, and 96.1%, respectively.>

The derivations of a novel approach for simultaneous recursive weight adaptation and node creation in multilayer backpropagation neural networks are presented. The method uses time and order update formulations in the orthogonal projection method to derive a recursive weight updating procedure for the training process of the neural network and a recursive node creation algorithm for weight adjustment of a layer with added nodes during the training process. The proposed approach allows optimal dynamic node creation in the sense that the mean-squared error is minimized for each new topology. The effectiveness of the algorithm is demonstrated on several benchmark problems (the multiplexer and the decoder problems) as well as a real world application for detection and classification of buried dielectric anomalies using a microwave sensor.

An approach is introduced for the recursive computation of the principal components of a vector stochastic process. The neurons of a single-layer perceptron are sequentially trained using a recursive least squares (RLS)-type algorithm to extract the principal components of the input process. The proof of the convergence of the weights at the nth neuron to the nth principal component, given that the previous (n-1) training steps have determined the first (n-1) principal components, is established. Simulation results are given to show the accuracy and speed of this algorithm in comparison with previous methods.>

A two-dimensional method which uses a full-plane image model to generate a more accurate filtered estimate of an image that has been corrupted by additive noise and full-plane blur is presented. Causality is maintained within the filtering process by using multiple concurrent block estimators. In addition, true state dynamics are preserved, resulting in an accurate Kalman gain matrix. Simulation results on a test image corrupted by additive white Gaussian noise are presented for various image models and compared to those of the previous block Kalman filtering methods.

The authors present the derivations of a novel approach for recursive weight adaptation and node creation in multilayer perceptron neural networks. The method uses time and order update formulations in the orthogonal projection method to derive a recursive weight updating procedure for the training process of the neural network and a recursive node creation algorithm for weight adjustment of a layer with added nodes after the training process. The proposed approach allows optimal dynamic node creation in the sense that the mean-squared error is minimized for each new architecture.>

A method for removing speckle from synthetic aperture radar (SAR) imagery by using 2-D adaptive block Kalman filtering is introduced. The image process is represented by an autoregressive model with a nonsymmetric half-plane (NSHP) region of support. New 2-D Kalman filtering equations are derived which taken into account not only the effect of speckles as multiplicative noise but also the effects of the additive receiver thermal noise and the blur. This method assumes local stationarity within a processing window, whereas the image can be assumed to be globally nonstationary. A recursive identification process using the stochastic Newton approach is also proposed which can be used on-line to estimate the filter parameters based upon the information within each new block of the image. Simulation results on several images are provided to indicate the effectiveness of the proposed method when used to remove the effects of speckle noise as well as those of the additive noise.>

A new approach for the learning process of multilayer perceptron neural networks using a recursive-least-squares-(RLS) type algorithm is proposed. The weights in the network are updated recursively upon the arrival of a new training sample. To determine the desired target in the hidden layers an analog of the back-propagation strategy used in the conventional learning algorithms is developed. This permits the application of the learning procedure to all the other lower layers. Simulation results on the 4-b parity checker problem are provided.>

The development of parallel and pipeline structures for real-time 2-D scalar and block recursive filtering is discussed. A skew-pipeline algorithm is developed for 2-D block processing. It can be implemented using a 2-D systolic array architecture. This structure can is also be used for other general applications where high-speed matrix-matrix multiplications should be performed on systolic arrays. For general 2-D recursive filters, a 2-D interleaved pipeline architecture which is more compatible to the nature of image scanning is introduced. Using this novel structure, the sampling period can reach a fraction of the time needed for one scalar addition operation and the delay is only a few samples.>

A 2-D model reduction using singular perturbation methodology is presented. The strong-weak coupling effects that lead to eigenvalue clustering of A/sub 1/ and A/sub 4/ matrices are utilized to derive structure transformations for fast-slow decomposition. This in conjunction with an aggregation procedure given allows the development of reduced-order 2-D slow and fast subsystems that capture the essential behavior of the full-order system as evidenced by close examination of the relevant characteristic polynomials.>

For a class of 2-D digital filters realized by a 2-D state-space structure, two different methods for computing the coefficients of their characteristic polynomial are introduced. The relative efficiencies of these methods are compared with that of the conventional method. An efficient method is then chosen which computes the coefficients with a minimum number of multiplications. The effectiveness of this method has been examined on a numerical example.

The paper considers error analysis of block implemented 2-D digital filters. An expression for bound on the roundoff error accumulation is derived using fixed-point arithmetic, and compared with the results obtained using ordinary 2-D difference equations.