Speech Emotion Recognition (SER) is the task of recognizing a speaker's emotional state from speech. SER plays a significant role in Human-Computer Interaction and psychological assessment. Several kinds of time-frequency representations like spectrograms, mel-frequency cepstrum coefficients (MFCCs), and mel-spectrograms are commonly used to develop an SER system. These representations use the Fast Fourier Transform (FFT) to convert the time domain signal to the frequency domain. However, the FFT has one fundamental limitation due to the uncertainty principle, which does not simultaneously allow a good resolution in both time and frequency domains. On the other hand, the multiresolution property of wavelets can provide a good localization in both time and frequency domains. Therefore, this article investigates the competency of the wavelet transforms for SER. We propose a Wavelet based Deep Emotion Recognition (WaDER) method using an autoencoder and 1D convolutional neural network (CNN) and long short-term memory (LSTM) networks. The autoencoder is used to perform the dimensionality reduction of the wavelet features then the latent space is used to classify the emotions using the 1D CNN-LSTM model. We conducted a Monte-Carlo K-fold validation using the Ryerson Audio-Visual Database of Emotional Speech and Song (RAVDESS) dataset. For speaker-dependent (SD) experiments, we achieved an unweighted accuracy (UA) of 81.45% and a weighted accuracy (WA) of 81.22%. The results of the experiments on the RAVDESS dataset show that the proposed method performs better than the state-of-the-art methods, which use other time-frequency representations.

Immunotherapy consists of assisting or augmenting the immune system to prevent or treat diseases and has shown great promise in improving upon the conventional methods of treating cancer, namely radiation therapy, chemotherapy, and surgery [2]. However, conducting preclinical studies in this field can take months to years and often require costly equipment and training to execute. Researchers have used in-silico methods to reduce the time to conduct these studies, computationally exploring many different experimental configurations in a fraction of the time needed in a wet lab [1]. A common method used to describe the interaction of the immune system with tumor cells is through ordinary and partial differential equations of the evolution in the populations, but these top-down approaches often cannot capture the necessary detail required to understand the dynamics that operate at the cellular level [3]. Appropriately, studies that approach these systems from the bottom-up, particularly agent-based simulations, have shown success in capturing the dynamics of cell-to-cell interactions [4].

With the impending launch of numerous hyperspectral satellites, the number of multitemporal datasets will exponentially increase, greatly increasing our need to develop algorithms designed to handle this imagery. Classifying scenes over time is challenging due to factors such as illumination changes due to the time of year, vegetation phenology, and changes in phenology patterns due to differences in precipitation and temperature. In order to improve multitemporal classifications, this letter investigates multiple algorithms that transform training data with no labels to classify future scenes. Four classes of algorithms are investigated to execute the transformations: linear, affine, decision tree, and random forest$(\mathcal {L}, \mathcal {A}, \mathcal {D}$, and$\mathcal {R}$). Experiments were conducted with hyperspectral airborne data collected three times per year for three consecutive years over Southern California. Classification accuracies with and without the transformations were compared. The mean improvements in accuracy achieved using the transformations were significant: 38%, 38%, 20%, and 22% for the four algorithm classes,$(\mathcal {L}, \mathcal {A}, \mathcal {D}$, and$\mathcal {R}$), respectively. These transformation algorithms present an opportunity for improving multitemporal classifications, even for challenging scenes, with simple calculations that are less computationally complex than existing algorithms. This approach represents another step to fully leveraging the unique and powerful temporal capabilities that will become available soon through upcoming or recently launched hyperspectral satellites.

In remote sensing, it is often challenging to acquire or collect a large data set that is accurately labeled. This difficulty is usually due to several issues, including but not limited to the study site’s spatial area and accessibility, errors in the global positioning system (GPS), and mixed pixels caused by an image’s spatial resolution. We propose an approach, with two variations, that estimates multiple-target signatures from training samples with imprecise labels: multitarget multiple-instance adaptive cosine estimator (MTMI-ACE) and multitarget multiple-instance spectral match filter (MTMI-SMF). The proposed methods address the abovementioned problems by directly considering the multiple-instance, imprecisely labeled data set. They learn a dictionary of target signatures that optimizes detection against a background using the adaptive cosine estimator (ACE) and spectral match filter (SMF). Experiments were conducted to test the proposed algorithms using a simulated hyperspectral data set, the MUUFL Gulfport hyperspectral data set collected over the University of Southern Mississippi–Gulfpark Campus, and the Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) hyperspectral data set collected over Santa Barbara County, CA, USA. Both simulated and real hyperspectral target detection experiments show that the proposed algorithms are effective at learning target signatures and performing target detection.

Combining a hyperspectral (HS) image and a multispectral (MS) image - an example of image fusion - can result in a spatially and spectrally high-resolution image. Despite the plethora of fusion algorithms in remote sensing, a necessary prerequisite, namely registration, is mostly ignored. This limits their application to well-registered images from the same source. In this article, we propose and validate an integrated registration and fusion approach (code available at https://github.com/zhouyuanzxcv/Hyperspectral). The registration algorithm minimizes a least-squares (LSQ) objective function with the point spread function (PSF) incorporated together with a nonrigid freeform transformation applied to the HS image and a rigid transformation applied to the MS image. It can handle images with significant scale differences and spatial distortion. The fusion algorithm takes the full high-resolution HS image as an unknown in the objective function. Assuming that the pixels lie on a low-dimensional manifold invariant to local linear transformations from spectral degradation, the fusion optimization problem leads to a closed-form solution. The method was validated on the Pavia University, Salton Sea, and the Mississippi Gulfport datasets. When the proposed registration algorithm is compared to its rigid variant and two mutual information-based methods, it has the best accuracy for both the nonrigid simulated dataset and the real dataset, with an average error less than 0.15 pixels for nonrigid distortion of maximum 1 HS pixel. When the fusion algorithm is compared with current state-of-the-art algorithms, it has the best performance on images with registration errors as well as on simulations that do not consider registration effects.

Hyperspectral unmixing while considering endmember variability is usually performed by the normal compositional model, where the endmembers for each pixel are assumed to be sampled from unimodal Gaussian distributions. However, in real applications, the distribution of a material is often not Gaussian. In this paper, we use Gaussian mixture models (GMM) to represent endmember variability. We show, given the GMM starting premise, that the distribution of the mixed pixel (under the linear mixing model) is also a GMM (and this is shown from two perspectives). The first perspective originates from random variable transformations and gives a conditional density function of the pixels given the abundances and GMM parameters. With proper smoothness and sparsity prior constraints on the abundances, the conditional density function leads to a standard maximum a posteriori (MAP ) problem which can be solved using generalized expectation maximization. The second perspective originates from marginalizing over the endmembers in the GMM, which provides us with a foundation to solve for the endmembers at each pixel. Hence, compared to the other distribution based methods, our model can not only estimate the abundances and distribution parameters, but also the distinct endmember set for each pixel. We tested the proposed GMM on several synthetic and real datasets, and showed its potential by comparing it to current popular methods.

We propose LOCO-CV-GP, a method for cross-validating Gaussian process (GP) methods in a leave-one-crown-out (LOCO) manner, when the GP method is applied on hyperspectral data from tree crowns. The fact that spectra within a crown are correlated [1] needs to be taken into consideration when working with airborne HS tree spectra. The experiments are conducted on OSBS2014 dataset to cross-validate OGP, a soft one-class classification method based on GP [2, 3]. Specifically, LOCO is justified by comparing LOCO-CV-GP to CV-GP where signals within a crown is split between cross-validation (CV) folds. Even though the proposed method was motivated by HS tree classification, it can be applied to any data where signal correlation happen within known partitions of the training signals.

Several techniques exist for dealing with spectral variability in hyperspectral unmixing, such as multiple endmember spectral mixture analysis (MESMA) or compositional models. These algorithms are computationally very involved, and often cannot be executed on problems of reasonable size. In this work, we present a new algorithm for solving the unmixing problem when spectral variability is present. The algorithm uses a library-based approach to describe the variability present in each class, and executes an alternating optimization with respect to these libraries. The optimization problem itself is constructed as an angle minimization problem by exploiting the geometrical interpretation of the unmixing problem. This results in an algorithm which yields almost identical results as MESMA, but is computationally much more favorable.

In many applications data classification may be hindered by the existence of multiple contexts that produce an input sample. To alleviate the problems associated with multiple contexts, context‐based classification is a process that uses different classifiers depending on a measure of the context. Context‐based classifiers offer the promise of increasing performance by allowing classifiers to become experts at classifying input samples of certain types, rather than trying to force single classifiers to perform well on all possible inputs. This study introduces a novel mixture of experts (ME) model, the mixture of hidden Markov model experts, for context‐based classification of samples that are variable length sequences; and derives the update equations for a single probabilistic model that to learn the experts and a gate that connects the experts. The model has a similar high‐level structure to the ME model but has the novelty that the gates and the experts are HMMs and the input data are sequences. Experimental results are presented on three datasets including one for landmine detection. Detailed analysis of the model is provided; which, over multiple runs and cross‐validation experiments, show superior results over the compared algorithms.

With the greater availability of imaging spectrometer data, vegetation species classification in the presence of outlier and ambiguous spectra is an increasingly important and poorly addressed problem. At large scales, assuming that all test spectra are from one of the training classes is unrealistic. An attractive resolution of these problems is the possibility theory, which is an axiomatic system, like probability theory, that represents uncertain labels of outliers and ambiguities more flexibly. In this letter, two popular probabilistic classification algorithms, namely, the support vector machine (SVM) with Platt scaling (SVM-Platt) and the Gaussian process (GP) classifier (GPC), are evaluated and compared to a novel one-class GP (OCGP) possibilistic classifier. OCGP, unlike one-class classification with GP, another GP-based one-class classifier, finds the hyperparameters automatically. Experiments were conducted with two data sets. The OCGP outperformed SVM-Platt and GPC when tested with outlier spectra.

An extension to the sparsity promoting iterated constrained endmember (SPICE) algorithm, named as SPICEE, has been presented. In ICE and SPICE, endmembers are estimated using a pseudoinverse method, which may generate endmembers that are not physically possible when representing normalized reflectance spectra. Although this problem can be alleviated by increasing the regularization, too much regularization leads to finding erroneous endmembers. To solve these problems, in this letter, a quadratic optimization solution is proposed that constrains the endmembers to have values between zero and one. The results on three data sets indicate that when regularization is large enough, SPICE and SPICEE generate similar answers; and when regularization is small to none, SPICEE stays more robust. In doing so, besides generating realistic endmembers, SPICEE helps in decreasing the effort necessary for fine-tuning the regularization parameter.

In hyperspectral unmixing applications, one typically assumes that a single spectrum exists for every endmember. In many scenarios, this is not the case, and one requires a set or a distribution of spectra to represent an endmember or class. This inherent spectral variability can pose severe difficulties in classical unmixing approaches. In this paper, we present a new algorithm for dealing with endmember variability in spectral unmixing, based on the geometrical interpretation of the resulting unmixing problem, and an alternating optimization approach. This alternating-angle-minimization algorithm uses sets of spectra to represent the variability present in each class and attempts to identify the subset of endmembers which produce the smallest reconstruction error. The algorithm is analogous to the popular multiple endmember spectral mixture analysis technique but has a much more favorable computational complexity while producing similar results. We illustrate the algorithm on several artificial and real data sets and compare with several other recent techniques for dealing with endmember variability.

The normal compositional model (NCM) has been extensively used in hyperspectral unmixing. However, previous research has mostly focused on estimation of endmembers and/or their variability, based on the assumption that the pixels are independent random variables. In this paper, we show that this assumption does not hold if all the pixels are generated by a fixed endmember set. This introduces another concept, endmember uncertainty, which is related to whether the pixels fit into the endmember simplex. To further develop this idea, we derive the NCM from the ground up without the pixel independence assumption, along with (i) using different noise levels at different wavelengths and (ii) using a spatial and sparsity promoting prior for the abundances. The resulting new formulation is called the spatial compositional model (SCM) to better differentiate it from the NCM. The SCM maximum a posteriori (MAP) objective leads to an optimization problem featuring noise weighted least-squares minimization for unmixing. The problem is solved by projected gradient descent, resulting in an algorithm that estimates endmembers, abundances, noise variances, and endmember uncertainty simultaneously. We compared SCM with current state-of-the-art algorithms on synthetic and real images. The results show that SCM can in the main provide more accurate endmembers and abundances. Moreover, the estimated uncertainty can serve as a prediction of endmember error under certain conditions.

We present a Bayesian probabilistic model and inference algorithm for fuzzy clustering that provides expanded capabilities over the traditional Fuzzy C-Means approach. Additionally, we extend the Bayesian Fuzzy Clustering model to handle a variable number of clusters and present a particle filter inference technique to estimate the model parameters including the number of clusters. We show results on synthetic and real data and compare with other approaches.

A novel multiple-instance hidden Markov model (MI-HMM) is introduced for classification of time-series data, and its training is developed using stochastic expectation maximization. The MI-HMM provides a single statistical form to learn the parameters of an HMM in a multiple-instance learning framework without introducing any additional parameters. The efficacy of the model is shown both on synthetic data and on a real landmine data set. Experiments on both the synthetic data and the landmine data set show that an MI-HMM can achieve statistically significant performance gains when compared with the best existing HMM for the landmine detection problem, eliminate the ad hoc approaches in training set selection, and introduce a principled way to work with ambiguous time-series data.

A new class of subpixel target detection algorithms that use a local structured background model is introduced. Our approach, referred to as Context Dependent Target Detectors, extends existing structured detectors to multiple contexts. It is based on a robust context dependent spectral unmixing algorithm that uses multiple linear models to take into account the background variability. The claim is that robust context dependent unmixing provides a better description of the background with minimum target leakage, compared to global unmixing, and hence results in a better target-background separation. The approach is evaluated using the Orthogonal Subspace Projection (OSP) detector, the Adaptive Matched Subspace Detector (AMSD) and the Hybrid Subspace Detector (HSD). Experimental results show that our approach outperforms the baseline detectors (that use one global model for background).

A robust hyper spectral unmixing algorithm that finds multiple sets of end members is introduced. The algorithm, called Robust Context Dependent Spectral Unmixing (RCDSU), combines the advantages of context dependent unmixing and robust clustering. RCDSU adapts the unmixing to different regions, or contexts, of the spectral space. It combines fuzzy and possibilistic clustering and linear unmixing to learn multiple contexts and the optimal end members and abundances for each context. RCDSU uses fuzzy membership functions to partition the data, and possibilistic membership functions to identify noise and outliers. An extension of RCDSU to deal with the case of an unknown number of contexts is also proposed. The performance of the proposed work is evaluated using simulated and real hyper spectral data. The experiments show that the proposed methods can handle noisy data and identify an "optimal" number of contexts and appropriate end members within each context.

In this manuscript, we investigate which similarity measure discriminates the most between reflectance spectra of two sets of labeled hyperspectral pixels. We find preliminary evidence that shared nearest neighbor, a secondary similarity measure unknown to hyperspectral community, relatively helps the separability of a primary similarity measure such as radial basis function.

We present a new model for nonlinear spectral mixing observed in hyperspectral imagery and demonstrate how this model can be used for unmixing and obtaining abundance maps. The model is based on the idea that a single pixel can contain several spatially segregated areas containing different mineral mixtures and fuses Hapke's radiative transfer model for intimate mineral mixtures with the traditional linear mixing model. The resulting model allows great flexibility for generating spectra, provides abundance coefficients in terms of total relative ground cover for each endmember, and can be reduced to several other nonlinear mixing models by an appropriate choice of the parameters. Experiments on laboratory mineral mixtures and real hyperspectral imagery show reduced reconstruction errors and more accurate abundance coefficients compared with the linear mixing model or the recently introduced multimixture pixel model. Moreover, the reconstruction error improvement can be used as a per-pixel measure of the size of the intimate mixing component.

A new algorithm for subpixel target detection in hyperspectral imagery is proposed which uses the PFCM-FLICM-PCE algorithm to model and estimate the parameters of the image background. This method uses the piece-wise convex mixing model with spatial-spectral constraints, and uses possibilistic and fuzzy clustering techniques to find the piece-wise convex regions and robustly estimate the parameters. A method for integrating the elevation measurements of a co-registered LiDAR sensor is also proposed. The performance of the proposed methods is demonstrated on a real-world dataset with emplaced detection targets.

A hyperspectral endmember detection and spectral unmixing algorithm that finds multiple sets of endmembers is presented. Hyperspectral data are often nonconvex. The Piecewise Convex Multiple-Model Endmember Detection algorithm accounts for this using a piecewise convex model. Multiple sets of endmembers and abundances are found using an iterative fuzzy clustering and spectral unmixing method. The results indicate that the piecewise convex representation estimates endmembers that better represent hyperspectral imagery composed of multiple regions where each region is represented with a distinct set of endmembers.

A Metropolis-within-Gibbs sampler for piecewise convex hyperspectral unmixing and endmember extraction is presented. The standard linear mixing model used for hyperspectral unmixing assumes that hyperspectral data reside in a single convex region. However, hyperspectral data are often nonconvex. Furthermore, in standard endmember extraction and unmixing methods, endmembers are generally represented as a single point in the high-dimensional space. However, the spectral signature for a material varies as a function of the inherent variability of the material and environmental conditions. Therefore, it is more appropriate to represent each endmember as a full distribution and use this information during spectral unmixing. The proposed method searches for several sets of endmember distributions. By using several sets of endmember distributions, a piecewise convex mixing model is applied, and given this model, the proposed method performs spectral unmixing and endmember estimation given this nonlinear representation of the data. Each set represents a random simplex. The vertices of the random simplex are modeled by the endmember distributions. The hyperspectral data are partitioned into sets associated with each of the extracted sets of endmember distributions using a Dirichlet process prior. The Dirichlet process prior also estimates the number of sets. Thus, the Metropolis-within-Gibbs sampler partitions the data into convex regions, estimates the required number of convex regions, and estimates endmember distributions and abundance values for all convex regions. Results are presented on real hyperspectral and simulated data that indicate the ability of the method to effectively estimate endmember distributions and the number of sets of endmember distributions.

A hyperspectral endmember detection and spectral unmixing algorithm that finds multiple sets of endmembers is presented. This algorithm, the Piece-wise Convex Multiple Model Endmember Detection (P-COMMEND) algorithm, models a hyperspectral image using a piece-wise convex representation. By using a piece-wise convex representation, non-convex hyperspectral data are more accurately characterized. For example, the well-known Indian Pines hyperspectral image is used as an example of a piece-wise convex collection of pixels. The convex regions, weights, endmembers and abundances are found using an iterative fuzzy clustering method. Results indicate that the piece-wise convex representation provides endmembers that better represent hyperspectral data sets over methods that use a single convex region.

This paper introduces a novel mixture of experts model, the Mixture of Hidden Markov Model Experts (MHMME). This model is designed to perform context-based classification of samples that are variable length sequences. The contexts are determined by the gates and the classifiers are determined by the experts. The gates and the experts are learned simultaneously using a single probabilistic model. Experimental results on landmine dataset show that MHMME significantly outperforms the HMM-based and ME-based models.

A compressive sensing framework is described for hyperspectral imaging. It is based on the widely used linear mixing model,LMM, which represents hyperspectral pixels as convex combinations of small numbers of endmember (material) spectra. The coefficients of the endmembers for each pixel are called proportions. The endmembers and proportions are often the sought-after quantities; the full image is an intermediate representation used to calculate them. Here, a method for estimating proportions and endmembers directly from compressively sensed hyperspectral data based onLMMis shown. Consequently, proportions and endmembers can be calculated directly from compressively sensed data with no need to reconstruct full hyperspectral images. If spectral information is required, endmembers can be reconstructed using compressive sensing reconstruction algorithms. Furthermore, given known endmembers, the proportions of the associated materials can be measured directly using a compressive sensing imaging device. This device would produce a multiband image; the bands would directly represent the material proportions.

We propose a novel method for fusing different classifiers outputs. Our approach, called context extraction for local fusion with fuzzy integrals (CELF-FI), is a local approach that adapts a fuzzy integrals fusion method to different regions of the feature space. It is based on a novel objective function that combines context identification and multialgorithm fusion criteria into a joint objective function. This objective function consists of two terms: The first is designed to produce compact clusters, called contexts, and the second is designed to produce Sugeno measures for fuzzy integral fusion for each context. The terms are optimized simultaneously via alternating optimization. To test a new sample, first, its features (extracted by each algorithm) are used to assign it to each context with a fuzzy membership degree. Second, the sample confidence values (assigned by each algorithm) are fused within each context using the learned context fusion parameters. Then, the context-dependent partial confidence values are weighted by the membership degrees and averaged over all contexts to produce a final confidence value. We illustrate the performance of CELF using synthetic data, and we apply it to the problem of landmine detection using ground penetrating radar and wideband electromagnetic induction. Our extensive experiments have indicated that the proposed fusion approach outperforms all individual classifiers, the global fuzzy integral fusion method, and the basic local fusion with linear aggregation.

In this paper, we provide a comprehensive survey of the mixture of experts (ME). We discuss the fundamental models for regression and classification and also their training with the expectation-maximization algorithm. We follow the discussion with improvements to the ME model and focus particularly on the mixtures of Gaussian process experts. We provide a review of the literature for other training methods, such as the alternative localized ME training, and cover the variational learning of ME in detail. In addition, we describe the model selection literature which encompasses finding the optimum number of experts, as well as the depth of the tree. We present the advances in ME in the classification area and present some issues concerning the classification model. We list the statistical properties of ME, discuss how the model has been modified over the years, compare ME to some popular algorithms, and list several applications. We conclude our survey with future directions and provide a list of publicly available datasets and a list of publicly available software that implement ME. Finally, we provide examples for regression and classification. We believe that the study described in this paper will provide quick access to the relevant literature for researchers and practitioners who would like to improve or use ME, and that it will stimulate further studies in ME.

Imaging spectroscopy refers to methods for identifying materials in a scene using cameras that digitize light into hundreds of spectral bands. Each pixel in these images consists of vectors representing the amount of light reflected in the different spectral bands from the physical location corresponding to the pixel. Images of this type are called hyperspectral images. Hyperspectral image analysis differs from traditional image analysis in that, in addition to the spatial information inherent in an image, there is abundant spectral information at the pixel or sub-pixel level that can be used to identify materials in the scene. Spectral unmixing techniques attempt to identify the material spectra in a scene down to the sub-pixel level. In this paper, a piece-wise convex hyperspectral unmixing algorithm using both spatial and spectral image information is presented. The proposed method incorporates possibilistic and fuzzy clustering methods. The typicality and membership estimates from those methods can be combined with traditional material proportion estimates to produce more meaningful proportion estimates than obtained with previous spectral unmixing algorithms. An analysis of the utility of using all three estimates produce a better estimate is given using real hyperspectral imagery.

Multiple instance learning (MIL) is a widely researched learning paradigm that allows a machine learning algorithm to learn target concepts from data with uncertain class labels. The random set framework for multiple instance learning (RSF-MIL) makes use of the random set to learn in this scenario of uncertainty. Previous models used assumptions that imposed a disjunctive relationship between the simple concepts learned (which compose the target concept). In the following, a conjunctive formulation of RSF-MIL is proposed and investigated. Results illustrate the utility of the conjunctive and disjunctive formulations of RSF-MIL and the scenarios when each is applicable.

Multi-kernel learning has become a popular method to allow classification models greater flexibility in representing the relationships between data points. This approach has evolved into localized multi-kernel learning, which creates classification models that have the ability to adapt to a multi-scale feature-space. The advantages of such an approach are often hampered by additional parameters and hyper-parameters involved in creating this model, not to mention the greater likelihood of over-training. Additionally, existing methods to create a localized multi-kernel classifier rely on partitioning the feature-space, followed by applying a multi-kernel to the partitioned data points. We introduce a Bayesian approach to the localized multi-kernel machine. The new model is shown to provide greater classification abilities by learning the local scales of the feature-space without the need to partition the data. Also, the Bayesian formulation helps the model to be resistant to over-training. We demonstrate the models effectiveness on two landmine detection datasets, each from a different sensor type.

Multiple-instance learning (MIL) is a learning paradigm used for learning a target concept in the presence of noise or with an uncertainty in target information including class labels. Due to the difficult situations in which hyperspectral images (HSIs) are collected, research in this area is extremely relevant and directly applicable. In the following, an MIL framework is proposed for target spectra learning for HSI analysis. MIL techniques are compared to their non-MIL counterparts (standard machine learning techniques). Experimental results indicate that MIL can learn target spectra with a lack of target information and, furthermore, result in improved classifiers.

We propose a multistream discrete hidden Markov model (DHMM) framework and apply it to the problem of land-mine detection using ground-penetrating radar (GPR). We hypothesize that each signature (mine or nonmine) can be characterized better by multiple synchronous sequences representing features that capture different environments and different radar characteristics. This paper is motivated by the fact that mines and clutter objects can have different characteristics depending on the mine type, soil and weather conditions, and burial depth. Thus, ideally different sets of specialized feature extraction mechanisms may be needed to achieve high detection and low false alarm rates. In order to fuse the different modalities, a multistream DHMM that includes a stream relevance weighting component is developed. The relevance weight of each stream depends on the symbols and the states. We reformulate the Baum-Welch and the minimum classification error/gradient probabilistic descent learning algorithms to include stream relevance weights and partial state probabilities. We generalize their objective functions and derive the necessary conditions to update all model parameters simultaneously. The results on a synthetic data set and a collection of GPR signatures show that the proposed multistream DHMM framework outperforms the basic single-stream DHMM where all the streams are treated equally important.

We propose a novel method for fusing different classifiers outputs. Our approach, called Context Extraction for Local Fusion with Fuzzy Integrals (CELF-FI), is a local approach that adapts fuzzy integrals fusion method to different regions of the feature space. It is based on a novel objective function that combines context identification and multi-algorithm fusion criteria into a joint objective function. This objective function is defined and optimized to produce contexts as compact clusters via unsupervised clustering. Optimization of the objective function also provide an optimal Sugeno measure within each context. Our initial experiments have indicated that the proposed fusion approach outperforms all individual classifiers, the global fuzzy integral fusion method, and the basic local fusion with linear aggregation.

Fuzzy c-means (FCM) is a simple but powerful clustering method using the concept of fuzzy sets that has been proved to be useful in many areas. There are, however, several well known problems with FCM, such as sensitivity to initialization, sensitivity to outliers, and limitation to convex clusters. In this paper, global fuzzy c-means (G-FCM) and kernel fuzzy c-means (K-FCM) are combined and extended to form a non-linear variant of G-FCM, called kernelized global fuzzy c-means (KG-FCM). G-FCM is a variant of FCM that uses an incremental seed selection method and is effective in alleviating sensitivity to initialization. There are several approaches to reduce the influence of noise and properly partition non-convex clusters, and K-FCM is one. K-FCM is used in this paper because it can easily be extended with different kernels, which provide sufficient flexibility to allow for resolution of the shortcomings of FCM. By combining G-FCM and K-FCM, KG-FCM can resolve the shortcomings mentioned above. The usefulness of the proposed method is demonstrated by experiments using artificial and real world data sets.

Context extraction for local fusion (CELF) is a local approach that combines multiple classifier outputs with the help of feature space information. CELF is based on an objective function that integrates context extraction and decision fusion. Context extraction divides the feature space into homogeneous regions; decision fusion combines multiple classifier outputs in each region or context. Although CELF is a generalization of previous fusion methods, it has some problems and noise sensitivity is one of them. The noise sensitivity problem comes from the fact that there are large number of parameters that should be optimized. Some variants of CELF were propose to mitigate the noise sensitivity problem and they showed better results than CELF. However, they also have their own shortcomings. In this paper, we proposed another variant of CELF to reduce the noise sensitivity and to overcome the shortcomings in the variants of CELF. The proposed method, CELF with regularization (CELF-R), uses regularization, which is a well-known method to reduce noise sensitivity by smoothing solution space. The noise robustness in CELF-R helps it to achieve better results than CELF and its variants. Experiments using landmine data sets also support this.

Multiple instance learning (MIL) is a recently researched technique used for learning a target concept in the presence of noise. Previously, a random set framework for multiple instance learning (RSF-MIL) was proposed; however, the proposed optimization strategy did not permit the harmonious optimization of model parameters. A cross entropy, based optimization strategy is proposed. Experimental results on synthetic examples, benchmark and landmine data sets illustrate the benefits of the proposed optimization strategy.

In this paper, we (1) provide a complete framework for classification using Variational Mixture of Experts (VME); (2) derive the variational lower bound; and (3) apply the method to landmine, or simply mine, detection and compare the results to the Mixtures of Experts trained with Expectation Maximization (EMME). VME has previously been used for regression and Waterhouse explained how to apply VME to classification (which we will call as VMEC). However, the steps to train the model were not made clear since the equations were applicable to vector valued parameters as opposed to matrices for each expert. Also, a variational lower bound was not provided. The variational lower bound provides an excellent stopping criterion that resists over-training. We demonstrate the efficacy of the method on real-world mine classification; in which, training robust mine classification algorithms is difficult because of the small number of samples per class. In our experiments VMEC consistently improved performance over EMME.

The Functions of Multiple Instances (FUMI) method for learning a target prototype from data points that are functions of target and non-target prototypes is introduced. In this paper, a specific case is considered where, given data points which are convex combinations of a target prototype and several non-target prototypes, the Convex-FUMI (C-FUMI) method learns the target and non-target patterns, the number of nontarget patterns, and determines the weights (or proportions) of all the prototypes for each data point. For this method, training data need only binary labels indicating whether the data contains or does not contain some proportion of the target prototype; the specific target weights for the training data are not needed. After learning the target prototype using the binary labeled training data, target detection is performed on test data. Results showing detection of the skin in hyper spectral imagery and sub-pixel target detection in simulated data are presented.

Multiple instance learning is a recently researched learning paradigm that allows a machine learning algorithm to learn target concepts with uncertainty in the class labels of training data. In the following, this approach is assessed for use in hyperspectral image analysis. Two leading MIL algorithms are used in a classification experiment and results are compared to a state-of-the-art context-based classifier. Results indicate that using a MIL based approach may improve learned target models and subsequently classification results.

We propose a discriminative method for combining heterogeneous sets of features for the continuous hidden Markov model classifier. We use a model level fusion approach and apply it to the problem of landmine detection using ground penetrating radar (GPR). We hypothesize that each signature (mine or non-mine) can be characterized better by multiple synchronous sequences that can capture different and complementary features. Our work is motivated by the fact that mines and clutter objects can have different characteristics depending on the mine type, soil and weather conditions, and burial depth. Thus, different sets of specialized feature extraction mechanisms, may be needed to achieve high detection and low false alarm rates. In order to fuse the different modalities, a multi-stream continuous HMM that includes a stream relevance weighting component is developed. In particular, we modify the probability density function that characterizes the standard continuous HMM to include state and component dependent stream relevance weights. We generalize the Minimum Classification Error (MCE) objective function to include stream relevance weights and derive the necessary conditions to update all model parameters simultaneously. Results on a large collection of GPR alarms show that the proposed model level fusion outperforms the baseline HMM when each feature is used independently and when both features are combined with equal weights.

The results from L1-Endmembers display the algorithm's stability and accuracy with increasing levels of noise. The algorithm was extremely stable in the number of endmembers when compared to the SPICE algorithm and the Virtual Dimensionality methods for estimating the number of endmembers. Furthermore, the results shown for this algorithm were generated with the same parameter set for all of the data sets, from two-dimensional data to 51-dimensional real hyperspectral data. This indicates L1-Endmembers may lack of sensitivity to parameter value settings. The L1-Endmembers algorithm requires several quadratic programming steps per iteration. These can be completed directly in quadratic programming software packages such as CPLEX and take advantage of any running time reductions the software packages provide. Investigations will be conducted into whether the specific form of this algorithm, particularly with respect to the constraints on the abundance values, can be used to reduce the running time.

This letter presents a simple and fast algorithm to analyze wideband electromagnetic induction data for subsurface targets. A well-known four-parameter model is differentiated, resulting in a two-parameter model. A fast lookup table is used to find parameters as opposed to nonlinear optimization. The proposed approach provides a computationally faster way to reproduce the results of state-of-the-art methods. A detailed mathematical analysis of the model is given that describes the advantages and limitations of the proposed method.

We present a novel method for fusing the results of multiple land mine detection algorithms which use different sensors, features, and different classification methods. The proposed multisensor/multialgorithm fusion method, which is called context-dependent fusion (CDF), is motivated by the fact that the relative performance of different sensors and algorithms can vary significantly depending on the mine type, geographical site, soil and weather conditions, and burial depth. CDF is a local approach that adapts the fusion method to different regions of the feature space. The training part of CDF has two components: context extraction and algorithm fusion. In context extraction, the features used by the different algorithms are combined and used to partition the feature space into groups of similar signatures, or contexts. The algorithm fusion component assigns a degree of worthiness to each detector in each context based on its relative performance within the context. To test a new alarm using CDF, each detection algorithm extracts its set of features and assigns a confidence value. Then, the features are used to identify the best context, and the degrees of worthiness of this context are used to fuse the individual confidence values. Results on large and diverse ground-penetrating radar and wideband electromagnetic data collections show that the proposed method can identify meaningful and coherent clusters and that different expert algorithms can be identified for the different contexts. Typically, the contexts correspond to groups of alarm signatures that share a subset of common features. Our extensive experiments have also indicated that CDF outperforms all individual detectors and the global fusion that uses the same method to assign aggregation weights.

A new hyperspectral endmember detection method that represents endmembers as distributions, autonomously partitions the input data set into several convex regions, and simultaneously determines endmember distributions (EDs) and proportion values for each convex region is presented. Spectral unmixing methods that treat endmembers as distributions or hyperspectral images as piecewise convex data sets have not been previously developed. Piecewise convex endmember (PCE) detection can be viewed in two parts. The first part, the ED detection algorithm, estimates a distribution for each endmember rather than estimating a single spectrum. By using EDs, PCE can incorporate an endmember's inherent spectral variation and the variation due to changing environmental conditions. ED uses a new sparsity-promoting polynomial prior while estimating abundance values. The second part of PCE partitions the input hyperspectral data set into convex regions and estimates EDs and proportions for each of these regions. The number of convex regions is determined autonomously using the Dirichlet process. PCE is effective at handling highly mixed hyperspectral images where all of the pixels in the scene contain mixtures of multiple endmembers. Furthermore, each convex region found by PCE conforms to the convex geometry model for hyperspectral imagery. This model requires that the proportions associated with a pixel be nonnegative and sum to one. Algorithm results on hyperspectral data indicate that PCE produces endmembers that represent the true ground-truth classes of the input data set. The algorithm can also effectively represent endmembers as distributions, thus incorporating an endmember's spectral variability.

We present a novel method for fusing different classifiers outputs. Our approach, called Context Extraction for Local Fusion with Feature Discrimination (CELF-FD), is a local approach that adapts the fusion method to different regions of the feature space. It is based on a novel objective function that combines context identification and multi-algorithm fusion criteria into a joint objective function. This objective function is defined and optimized to produce contexts as compact clusters in subspaces of the high-dimensional feature space via unsuper-vised clustering and feature discrimination. Optimization of the objective function also provide optimal fusion parameters for each context. Our initial experiments have indicated that the proposed fusion approach outperforms all individual classifiers and the global fusion method.

Support vector machine (SVM) is a theoretically well motivated algorithm developed from statistical learning theory, that have shown good performance in many fields. In spite of its success, it still suffers from a noise sensitivity problem. To relax this problem, the SVM was extended by the introduction of fuzzy memberships to the fuzzy SVM (FSVM). The FSVM also has been extended further in two ways: by adopting a different objective function with the help of domain-specific knowledge and by employing a different membership calculation method. In this paper, we propose a new membership calculation method, that belongs to the second group. It is different from previous ones in that it does not assume any simple data distribution and does not need any prior knowledge. The proposed method is based on reconstruction error, which measures the agreement between the overall data structure and a data point. Thus the reconstruction error can represent the degree of outlier-ness and help in achieving noise robustness. Experimental results with synthetic and real data sets also support this.

We propose a modified discrete HMM that handles multimodalities. We assume that the feature space is partitioned into subspaces generated by different sources of information. To combine these heteregoneous modalities we propose a multi-stream discrete HMM that assigns a relevance weight to each subspace. The relevance weights are set local and depend on the symbols and the states. In particular, we associate a partial probability with each symbol in each subspace. The overall observation state probability is then computed as an aggregation of the partial probabilities and their objective relevance weights based on a linear combination. The minimum classification error (MCE) objective based on the gradient probabilistic descent (GPD) optimization algorithm is reformulated to derive the update equations for the relevance weights and the partial state probabilities. The proposed approach is validated using synthetic and real data sets. The results are shown to outperform the baseline discrete HMM that treats all streams equally important.

We present a novel method for fusing the decisions of multiple classification algorithms which use different features, classification methods, and data sources. The proposed method, called context dependent fusion of multiple algorithms (CDF-MA) is motivated by the fact that the relative performance of different algorithms can vary significantly as the characteristics of the input data vary. The training part of CDF-MA has two main components: context extraction and algorithm fusion. In context extraction, the features used by the distinct algorithms are combined and clustered into groups of similar contexts. The algorithm fusion component embeds a variation of the MCE/GPD method to assigns an aggregation weight to each algorithm based on its loss function within each context. Results on real world data show that the proposed method can identify meaningful and coherent clusters, and outperform all individual classifiers and the global weighted average fusion method.

This paper addresses the problem of estimating the correct number of components in a Gaussian mixture given a sample data set. In particular, an extension of Gaussian-means (G-means) and projected Gaussian-means (PG-means) algorithms is proposed. All these methods are based on one-dimensional statistical hypothesis test. G-means and PG-means are wrapper algorithms of the k-means and expectation-maximization (EM) algorithms, respectively. Although G-means is a simple and fast algorithm, it does not perform well when clusters overlap since it is based on k-means. PG-means can handle overlapped clusters but requires more computation and sometimes fails to find the right number of clusters. In this paper, we propose an extension, called extended projected Gaussian means (XPG-means). XPG-means is a wrapper algorithm of possibilistic fuzzy C-means (PFCM) algorithm. XPG-means integrates the advantages of both algorithms while resolving some of the disadvantages involving overlapped clusters, noise, and computational complexity. More specifically, XPG-means handles overlapped clusters better than G-means because of the use of fuzzy clustering, handles noise better than both algorithms because it uses possibilitistic clustering. XPG-means is less computationally expensive than PG-means because it uses local hypothesis testing scheme used by G-means that is specific to Gaussians whereas PG-means uses a more general Kolmogorov-Smirnow test on Gaussian mixtures. In addition, XPG-means demonstrates less variance in estimating the number of components than either of the other algorithms.

Principal component analysis (PCA) is widely used for dimensionality reduction in pattern recognition. Although PCA has been applied in many areas successfully, it suffers from sensitivity to noise and is limited to linear principal components. The noise sensitivity problem comes from the least-squares measure used in PCA and the limitation to linear components originates from the fact that PCA uses an affine transform defined by eigenvectors of the covariance matrix and the mean of the data. In this paper, a robust kernel PCA method that extends Scholkopf et al.'s kernel PCA and uses fuzzy memberships is introduced to tackle the two problems simultaneously. We first propose an iterative method to find a robust covariance matrix called robust fuzzy PCA (RF-PCA). The RF-PCA is introduced to reduce the sensitivity to noise with the help of robust estimation technique. The RF-PCA method is then extended to a non-linear one, robust kernel fuzzy PCA (RKF-PCA), using kernels. Experimental results suggest that the proposed algorithm works well on artificial and real world data sets.

This paper describes an algorithm for land mine detection using sensor data generated by a ground-penetrating radar (GPR) system that uses edge histogram descriptors for feature extraction and a possibilisticK-nearest neighbors (K-NNs) rule for confidence assignment. The algorithm demonstrated the best performance among several high-performance algorithms in extensive testing on a large real-world datasets associated with the difficult problem of land mine detection. The superior performance of the algorithm is attributed to the use of the possibilisticK-NN algorithm, thereby providing important evidence supporting the use of possibilistic methods in real-world applications. The GPR produces a 3-D array of intensity values, representing a volume below the surface of the ground. First, a computationally inexpensive prescreening algorithm for anomaly detection is used to focus attention and identify candidate signatures that resemble mines. The identified regions of interest are processed further by a feature extraction algorithm to capture their salient features. We use translation-invariant features that are based on the local edge distribution of the 3-D GPR signatures. Specifically, each 3-D signature is divided into subsignatures, and the local edge distribution for each subsignature is represented by a histogram. Next, the training signatures are clustered to identify prototypes. The main idea is to identify few prototypes that can capture the variations of the signatures within each class. These variations could be due to different mine types, different soil conditions, different weather conditions, etc. Fuzzy memberships are assigned to these representatives to capture their degree of sharing among the mines and false alarm classes. Finally, a possibilisticK-NN-based rule is used to assign a confidence value to distinguish true detections from false alarms. The proposed algorithm is implemented and integrated within a complete land mine prototype system. It is trained, field-tested, evaluated, and compared using a large-scale cross-validation experiment that uses a diverse dataset acquired from four outdoor test sites at different geographic locations. This collection covers over 41 807 m2of ground and includes 1593 mine encounters.

In this paper, a new matching pursuits dissimilarity measure (MPDM) is presented that compares two signals using the information provided by their matching pursuits (MP) approximations, without requiring any prior domain knowledge. MPDM is a flexible and differentiable measure that can be used to perform shape-based comparisons and fuzzy clustering of very high-dimensional, possibly compressed, data. A novel prototype based classification algorithm, which is termed the computer aided minimization procedure (CAMP), is also proposed. The CAMP algorithm uses the MPDM with the competitive agglomeration (CA) fuzzy clustering algorithm to build reliable shape based prototypes for classification. MP is a well known sparse signal approximation technique, which is commonly used for video and image coding. The dictionary and coefficient information produced by MP has previously been used to define features to build discrimination and prototype based classifiers. However, existing MP based classification applications are quite problem domain specific, thus making their generalization to other problems quite difficult. The proposed CAMP algorithm is the first MP based classification system that requires no assumptions about the problem domain and builds a bridge between the MP and fuzzy clustering algorithms. Experimental results also show that the CAMP algorithm is more resilient to outliers in test data than the multilayer perceptron (MLP) and support-vector-machine (SVM) classifiers, as well as prototype-based classifiers using the Euclidean distance as their dissimilarity measure.

In remotely sensed hyperspectral imagery, many samples are collected on a given flight and many variable factors contribute to the distribution of samples. Various factors transform spectral responses causing them to appear differently in different contexts. We develop a method that infers context via spectra population distribution analysis. In this manner, feature space orientations of sets of spectral signatures are characterized using random set models. The models allow for the characterization of complex and irregular patterns in a feature space. The developed random set framework for context-based classification applies context-specific classifiers in an ensemblelike manner, and aggregates their decisions based on their contextual relevance to the spectra under test. Results indicate that the proposed method improves classification accuracy over similar classifiers, which make no use of contextual information, and performs well when compared to similar context-based approaches.

During the 1990s Ritter, introduced a new family of associative memories based on lattice algebra instead of linear algebra. These memories provide unlimited storage capacity, unlike linear-correlation-based models. The canonical lattice-based memories, however, are susceptible to noise in the initial input data. In this brief, we present novel methods of encoding and decoding lattice-based memories using two families of ordered weighted average (OWA) operators. The result is a greater robustness to distortion in the initial input data, and a greater understanding of the effect of the choice of encoding and decoding operators on the behavior of the system, with the tradeoff that the time complexity for encoding is increased.

In many scientific fields, data classification may be hindered by population correlated factors or hidden contexts. These factors greatly affect samplespsila values making it difficult for standard classification models to perform well on a consistent basis. A general random set model is presented for context-based classification. An implementation is provided based on Possibility Theory. The result is a robust classifier that can intrinsically identify hidden contexts and classify data accordingly. The random set model is compared to standard kNN and set-based kNN. Results from synthetic data illustrate the random set modelpsilas ability to consistently improve classification through context estimation.

This paper applies the OWA aggregation operator to hand-held GPR data to improve the detection of landmines. Data from a number of sweeps are collected when the hand-held detector is operating in discrimination mode. The energy density spectra of the GPR signal return from individual sweeps are estimated and two OWA aggregation operations are performed to select the good quality sweeps that will be used for landmine detection. Experimental results using the real data collected from two different test sites show that the OWA operations provide significant performance improvement for landmine detection at high probability of detection.

Matching pursuits is a well known technique for signal representation and has also been used as a feature extractor for some classification systems. However, applications that use matching pursuits (MP) algorithm in their feature extraction stage are quite problem domain specific, making their adaptation for other types of problems quite hard. In this paper we propose a matching pursuits based similarity measure that uses only the dictionary, coefficients and residual information provided by the MP algorithm while comparing two signals. Hence it is easily applicable to a variety of problems. We show that using the MP based similarity measure for competitive agglomerative fuzzy clustering leads to an interesting and novel update equation that combines the standard fuzzy prototype updating equation with a term involving the error between approximated signals and approximated prototypes. The potential value of the similarity measure is investigated using the fuzzy k-nearest prototype algorithm of Frigui for a two-class, signal classification problem. It is shown that the new similarity measure significantly outperforms the Euclidean distance.

A novel algorithm is introduced for learning fuzzy measures for Choquet integral-based information fusion. The new algorithm goes beyond previously published MCE-based approaches. It has the advantage that it is applicable to general measures, as opposed to only the Sugeno class of measures. In addition, the monotonicity constraints are handled easily with minimal time or storage requirements. Learning the fuzzy measure is framed as a maximum a posteriori (MAP) parameter learning problem. In order to maintain the constraints, this MAP problem is solved with a Gibbs sampler using an expectation maximization (EM) framework. For these reasons, the new algorithm is referred to as the MAP-EM MCE logistic LASSO algorithm. Results are given on synthetic and real data sets, the latter obtained from a landmine detection problem. Average reductions in false alarms of about 25% are achieved on the landmine detection problem and probabilities of detection in the interesting and meaningful range of 85%-95%.

Ensemble learning is a variational Bayesian method in which an intractable distribution is approximated by a lower-bound. Ensemble learning results in models with better generalization and is less likely to fall into local maxima than Baum-Welch learning. However, it does not fully make use of the statistics of training data. In this paper, we propose a prior-updating variant which combines the data-driven property of Baum-Welch learning and the generalization property of ensemble learning. First we present experimental results suggesting that ensemble learning is better than the Baum-Welch learning in the aspects mentioned above and then we introduce a prior-updating method using training data. The prior-updating ensemble learning performs better than Baum-Welch as well as the pure ensemble learning in experiments with an artificial and a real data set.

Sparse representations using overcomplete dictionaries are used in a variety of field such as pattern recognition and compression. However, the size of dictionary is usually a tradeoff between approximation speed and accuracy. In this paper we propose a novel technique called the Enhanced K-SVD algorithm (EK-SVD), which finds a dictionary of optimized size-for a given dataset, without compromising its approximation accuracy. EK-SVD improves the K-SVD dictionary learning algorithm by introducing an optimized dictionary size discovery feature to K-SVD. Optimizing strict sparsity and MSE constraints, it starts with a large number of dictionary elements and gradually prunes the under-utilized or similar-looking elements to produce a well-trained dictionary that has no redundant elements. Experimental results show the optimized dictionaries learned using EK-SVD give the same accuracy as dictionaries learned using the K-SVD algorithm while substantially reducing the dictionary size by 60%.

An endmember detection algorithm for hyperspectral imagery using the Dirichlet process to determine the number of endmembers in a hyperspectral image is described. This algorithm provides an estimate of endmember spectra, proportion maps, and the number of endmembers needed for a scene. Updates to the proportion vector for a pixel are sampled using the Dirichlet process. As opposed to previous methods that prune unnecessary endmembers, the proposed algorithm is initialized with one endmember and new endmembers are added through sampling as needed. Results are shown on a two-dimensional dataset and a simulated dataset using endmembers selected from an AVIRIS hyperspectral image.

In remotely sensed hyperspectral imagery, many samples are collected on a given flight and many variable factors contribute to the distribution of samples. Various environmental factors transform spectral responses causing them to appear differently in different environmental contexts. Previously, we developed and applied context-based classifiers to hyperspectral imagery, improving classification results. A new variant of our model is presented which incorporates Bayesian classifiers into the model. Classification results are compared to those of a standard Bayesian classifier to identify the direct benefits of context estimation.

We present a novel method for fusing the results of multiple landmine detection algorithms that use different types of features, different classification methods, and different sensors. The proposed fusion method, called context-dependent multi-sensor fusion (CDMSF) is motivated by the fact that the relative performance of different detectors can vary significantly depending on the sensor, mine type, geographical site, soil and weather conditions, and burial depth. The training part of CDMSF has two components: context extraction and algorithm fusion. In context extraction, the features used by the different algorithms are combined and used to partition the feature space into groups of similar signatures, or contexts. The algorithm fusion component assigns an aggregation weight to each detector in each context based on its relative performance within the context. Results on ground penetrating radar (GPR) and wideband electromagnetic induction (WEMI) data collections show that the proposed method can identify meaningful and coherent clusters and that different expert algorithms can be identified for the different contexts. Our initial experiments have also indicated that the context-dependent fusion outperforms all individual detectors.

A variety of algorithms are presented and employed in a hierarchical fashion to discriminate both Anti-Tank (AT) and Anti-Personnel (AP) landmines using data collected from Wideband Electro-Magnetic Induction (WEMI) and Ground Penetrating Radar (GPR) sensors mounted on a robotic platform. The two new algorithms for WEMI are based on the In-phase vs. Quadrature plot (the Argand diagram) of the complex measurement obtained at a single spatial location. The Angle Prototype Match method uses the sequence of angles as a feature vector. Prototypes are constructed from these feature vectors and used to assign mine confidence to a test sample. The Angle Model Based KNN method uses a two parameter model; where the parameters are fit to the In-phase and Quadrature data. For the GPR data, the Linear Prediction Processing and Spectral Features are calculated. All four features from WEMI and GPR are used in a Hierarchical Mixture of Experts model to increase the landmine detection rate. The EM algorithm is used to estimate the parameters of the hierarchical mixture. Instead of a two way mine/non-mine decision, the HME structure is trained to make a five way decision which aids in the detection of the low metal anti personnel mines.

This letter presents a simultaneous band selection and endmember detection algorithm for hyperspectral imagery. This algorithm is an extension of the sparsity promoting iterated constrained endmember (SPICE) algorithm. The extension adds spectral band weights and a sparsity promoting prior to the SPICE objective function to provide integrated band selection. In addition to solving for endmembers, the number of endmembers, and end- member fractional maps, this algorithm attempts to autonomously perform band selection and to determine the number of spectral bands required for a particular scene. Results are presented on a simulated data set and the AVIRIS Indian Pines data set. Experiments on the simulated data set show the ability to find the correct endmembers and abundance values. Experiments on the Indian Pines data set show strong classification accuracies in comparison to previously published results.

The discrete Choquet integral is a nonlinear transformation that integrates a real function with respect to a fuzzy measure. We show that the discrete Choquet integral defines a metric if and only if the corresponding measure satisfies certain monotonicity constraints, thereby completely characterizing the class of measures that induce a metric with the Choquet integral.

A novel algorithm for discriminative training of Choquet-integral-based fusion operators is described. Fusion is performed by Choquet integration of classifier outputs with respect to fuzzy measures. The fusion operators are determined by the parameters of fuzzy measures. These parameters are found by minimizing a minimum classification error (MCE) objective function. The minimization is performed with respect to a special class of measures, the Sugeno lambda-measures. An analytic expression is derived for the gradient of the Choquet integral with respect to the Sugeno lambda-measure. The new algorithm is applied to a landmine detection problem, and compared to previous techniques.

Ground penetrating radar (GPR)-based discrimination of landmines from clutter is known to be challenging due to the wide variability of possible clutter (e.g., rocks, roots, and general soil heterogeneity). This paper discusses the use of GPR frequency-domain spectral features to improve the detection of weak-scattering plastic mines and to reduce the number of false alarms resulting from clutter. The motivation for this approach comes from the fact that landmine targets and clutter objects often have different shapes and/or composition, yielding different energy density spectrum (EDS) that may be exploited for their discrimination (this information is also present in time-domain data, but in the frequency domain we can remove a phase if desired and can reveal better spatial characteristics and therefore often achieve greater robustness). This paper first applies the finite-difference time-domain (FDTD) modeling technique to establish the theoretical foundation. The method to generate EDS from GPR measurements is then described. The consistency of the frequency-domain features is examined through two different GPRs that have different spatial sampling rates and frequency bandwidths. Experimental results from several test sites, based on GPR data collected over buried mines and emplaced buried clutter objects, corroborate the theoretical development and the effectiveness of the proposed spectral feature to increase the accuracy of landmine detection and discrimination.

We develop a vegetation mapping method using long-wave hyperspectral imagery and apply it to landmine detection. The novel aspect of the method is that it makes use of emissivity skewness. The main purpose of vegetation detection for mine detection is to minimize false alarms. Vegetation, such as round bushes, may be mistaken as mines by mine detection algorithms, particularly in synthetic aperture radar (SAR) imagery. We employ an unsupervised vegetation detection algorithm that exploits statistics of emissivity spectra of vegetation in the long-wave infrared spectrum for identification. This information is incorporated into a Choquet integral-based fusion structure, which fuses detector outputs from hyperspectral imagery and SAR imagery. Vegetation mapping is shown to improve mine detection results over a variety of images and fusion models.

We apply a population-based classifier to Long Wave HyperSpectral Imagery (LWHSI) for the purposes of landmine detection. In LWHSI, there are many environmental factors that are correlated with groups of samples (pixels in an image) in sample populations (individual images). These factors greatly affect samples' values making it difficult for standard classification models to perform well on a consistent basis. Population-based classifiers capture information correlated with sample populations. We perform classification experiments over a range of LWHSI imagery and compare results between the population-based classifier and standard kNN. After analysis, we show that the use of population-correlated information in LWHSI greatly improves classification results and consistency.

The HSTAMIDS handheld landmine detection system has been used in a number of humanitarian demining activities. Existing algorithms used with this system to assist human in discrimination process do a better job than the operator alone. However, they are unable to model mine and clutter signatures completely, leading to inaccurate mine confidence assignment. This paper presents a matching pursuits (MP) based landmine detection system using an application-specific dictionary. Prototypes for mine and clutter classes are built using MP decomposition of class members. A fuzzy K-nearest neighbor rule is used to assign confidence values for mine and clutter discrimination. The proposed algorithm is demonstrated on data acquired from three different landmines test sites.

An extension of the Iterated Constrained Endmembers (ICE) that incorporates sparsity promoting priors to find the correct number of endmembers and simultaneously select informative spectral bands is presented. In addition to solving for endmembers and endmember fractional maps, this algorithm attempts to autonomously determine the number of endmembers required for a particular scene. The number of endmembers is found by adding a sparsity-promoting term to ICE’s objective function. Additionally, hyperspectral band selection is performed by incorporating weights associated with each hyperspectral band. A sparsity promoting term for the band weights is added to the objective function to perform band selection.

An extension of the iterated constrained endmember (ICE) algorithm that incorporates sparsity-promoting priors to find the correct number of endmembers is presented. In addition to solving for endmembers and endmember fractional maps, this algorithm attempts to autonomously determine the number of endmembers that are required for a particular scene. The number of endmembers is found by adding a sparsity-promoting term to ICE's objective function.

A common approach to training neural network classifiers in a supervised learning setting is to minimize the mean-square error (mse) between the network output for each labeled training sample and some desired output. In the context of landmine detection and discrimination, although the performance of an algorithm is correlated with the mse, it is ultimately evaluated by using receiver operating characteristic (ROC) curves. In general, the larger the area under the ROC curve (AUC), the better. We present a new method for maximizing the AUC. Desirable properties of the proposed algorithm are derived and discussed that differentiate it from previously proposed algorithms. A hypothesis test is used to compare the proposed algorithm to an existing algorithm. The false alarm rate achieved by the proposed algorithm is found to be less than that of the existing algorithm with 95% confidence

There has been significant amount of study on the use of ground-penetrating radar (GPR) for land-mine detection. This paper presents our analysis of GPR data collected at a U.S. Army test site using a new approach based on frequency subband processing. In this approach, from the radar data that have over 2.5 GHz of bandwidth, we compute separate radar images using the one wide (2 GHz) and four narrow (0.6 GHz) frequency subbands. The results indicate that signals for different frequency subbands are significantly different and give very different performance in land-mine detection. In addition, we also examine a number of features extracted from the GPR data, including magnitude and local-contrast features, ratio between copolarization and cross-polarization signals, and features obtained using polarimetric decomposition. Feature selection procedures are employed to find subsets of features that improve detection performance when combined. Results of land-mine detection, including performance on blind test lanes, are presented

A variety of algorithms for the detection of landmines and discrimination between landmines and clutter objects have been presented. We discuss four quite different approaches in using data collected by a vehicle-mounted ground-penetrating radar sensor to detect landmines and distinguish them from clutter objects. One uses edge features in a hidden Markov model; the second uses geometric features in a feed-forward order-weighted average network; the third employs spectral features as its basis; and the fourth clusters edge histograms. We present the results of a large-scale cross-validation evaluation that uses a diverse set of data collected over 41 807.57$\hbox{m}^{2}$of ground, including 1593 mine encounters. Finally, we discuss the results of that ranking and what one can conclude concerning the performance of these four algorithms in various settings.

This paper describes an algorithm for land mine detection using sensor data generated by a ground penetrating radar (GPR) system. The GPR produces a 3-D array of intensity values, representing a volume below the surface of the ground. First, a computationally inexpensive pre-screening algorithm is used to focus attention and identify regions with subsurface anomalies. The identified regions of interest are then processed by a feature extraction algorithm to capture their salient features. We use translation invariant features that are based on the local edge distribution of the 3-D GPR signatures. Finally, a fuzzy K-nearest neighbor rule is used to assign a confidence value to distinguish true detections from false alarms. The proposed algorithm is applied to data acquired from three outdoor test sites at different geographic locations.

In this work, we propose a hidden Markov model (HMM), called the linguistic HMM, suitable for processing sequences of fuzzy vectors. A fuzzy vector B is an n-tuple of fuzzy numbers. Since fuzzy numbers are often associated with linguistic terms, such as "small," "medium," etc., a fuzzy vector can also be called a linguistic vector. Similarly, an HMM that processes linguistic vectors can be called a "linguistic HMM." The derivation of the linguistic HMM (LHMM) from the continuous HMM is performed using the extension principle and the decomposition theorem. We prove that a LHMM behaves in the same fashion as the CHMM in the degenerate linguistic case when the fuzzy numbers are singletons (real numbers). We also provide an example where an LHMM was used for the recognition of a play (pick-and-shoot) during a basketball game. The positions of the players were described using spatial fuzzy relations. For the recognition experiment, we generated two sets of 100 sequences containing pick-and-shoot and non pick-and-shoot sequences, respectively. The LHMM results obtained for the fuzzy sequences were compared to the CHMM results obtained on a crisp version of the same sequences. The results obtained showed that the fuzzy spatial relations together with the LHMM provide a better description of the movement than the CHMM.

This paper introduces a system for landmine detection using the sensor data generated by a ground penetrating radar (GPR). The GPR produces a three-dimensional array of intensity values, representing a volume below the surface of the ground. First, a constant false alarm rate (CFAR) detector is used to focus the attention and identify the candidates that resemble mines. Next, we apply a feature extraction algorithm based on projecting the data onto the dominant eigenvectors in the training data. The training signatures are then clustered to identify a few representatives, and a fuzzy k-nearest neighbor rule is used to distinguish true detections from false alarms.

A hit-miss transform is defined using Choquet integrals with respect to random sets. In this context, random sets represent random shapes defined on the plane. Random sets are characterised by their capacity functional. Capacity functional is fuzzy measures. Choquet integrals with respect to random sets are interpreted as stochastic morphological operations. Specifically, the integrals represent the average probability that sets either intersect or are contained in the random sets. Formulas are defined for random erosions and dilations of disks and annuli with Gaussian radii. Applications to landmine detection are given.

Entropy terms are analyzed for use in a heterogeneous network involving ordered weighted averaging (OWA) operators and feed-forward neural networks, referred to as FOWA networks. The OWA operators are used to aggregate feature information and were previously used successfully in a landmine application. Weight vectors of OWA operators are special cases of Choquet capacities for which the Shannon entropy has been defined. One can also consider the quadratic Renyi entropy for OWA weight vectors. A term is added to the objective function of the FOWA network involving these entropy terms. An analysis of the derivatives used in gradient descent is conducted.

We discuss the application of Choquet integrals to multi-algorithm and multi-sensor fusion in landmine detection. Choquet integrals are defined. Specific classes of measures, the full and Sugeno measures, are described. Full measures are optimized via quadratic programming. A steepest descent algorithm for optimizing Sugeno measures is derived by applying implicit differentiation. Multiple detection algorithms are applied to hyper-spectral and synthetic aperture radar imagery. In addition, a LWIR vegetation index is computed using statistics of apparent emissivity. The detection algorithms are combined using an OR operator and Choquet integrals with respect to full and Sugeno measures. The Choquet integral with respect to the full measure achieves lower false alarm rates

Landmine detection is an important and yet challenging problem remains to be solved. Ground penetrating radar (GPR) is an effective sensor to detect landmines that are made of plastic or have low metal content. Most GPR signal processing algorithms apply processing in the time (depth) domain. This work proposes to use the frequency domain features from the GPR signal to improve the detection of weak plastic mines and to reduce the number false alarms due to clutter objects. The motivation comes from the fact that the energy density spectrum may be different between mine targets and clutter objects, although both may have strong GPR signal return in the time domain. Experimental results based on clutter lane data collected at a test site corroborate the effectiveness of the proposed spectral features to increase the accuracy for landmine detection.

An approach to detecting landmines using ground-penetrating radar (GPR) based on feature-based rules, order statistics, and adaptive whitening (FROSAW) is described. FROSAW relies on independent adaptation of whitening statistics in different depths and combining feature-based methods with anomaly detection using rules. Constant false alarm rate (CFAR) detectors are used for anomaly detection on the depth-dependent adaptively whitened data. A single CFAR confidence measure is computed via a function of order statistics. Anomalies are detected at locations with high CFAR confidence. Depth-dependent features are computed on regions containing anomalies. Rules based on the features are used to reject alarms that do not exhibit mine-like properties. The utility of combining the CFAR and feature-based methods is evaluated. The algorithms and analysis are applied to data acquired from tens of thousands of square meters from several outdoor test sites with a state-of-the-art array of GPR sensors.

Signal processing algorithms for hand-held mine detection sensors are described. The goals of the algorithms are to provide alarms to a human operator indicating the likelihood of the presence of a buried mine. Two modes of operations are considered: search mode and discrimination mode. Search mode generates an initial detection at a suspected location and discrimination mode confirms that the suspected location contains a land mine. Search mode requires that the signal processing algorithm generate a detection confidence value immediately at the current sample location and no delay in producing an alarm confidence is tolerable. Search mode detection has a high false-alarm rate. Discrimination mode allows the operator to interrogate the entire suspected location to eliminate false alarms. It does not require that the signal processing algorithm produce an alarm confidence immediately for the current sample location, but rather allows the system to process all the data acquired over the region before producing an alarm. This paper proposes discrimination mode processing algorithms for metal detectors (MDs), or electromagnetic induction sensors (EMIs), ground-penetrating radars (GPRs), and their fusion. The MD discrimination mode algorithm employs a model-based approach and uses the target model parameters to discriminate between mines and clutter objects. The GPR discrimination mode algorithm uses the consistency of detection as well as the shape of the detection peaks over several sweeps to improve the discrimination accuracy. The performances of the proposed algorithms were examined on a dataset collected at a government test site, and performance was compared with baseline techniques. Experimental results showed that the proposed method can reduce the probability of false alarm by as much as 70% at a 100% correct detection rate and performed comparable to the best human operator on a blind test with data collected at approximately 1000 locations.

This paper introduces a new system for real-time land mine detection using sensor data generated by a Ground Penetrating Radar (GPR). The GPR produces a three-dimensional array of intensity values, representing a volume below the surface of the ground. Features are computed from this array and two types of membership degrees are assigned to each location. A fuzzy membership value provides a degree of belongingness of a given observation in the classes of mines, false alarms, and background, while a possibilistic membership value provides a degree of typicality. Both membership degrees are combined using simple rules to assign a confidence value. The parameters of the membership functions are obtained by clustering the training data and using the statistics of each partition. Our preliminary results show that the proposed approach is simple, efficient, and yet, yields results comparable to more complex detection systems.

An interpretation of continuous Choquet integrals with respect to random sets is given in the context of image filtering and shape detection. In this context, random sets represent random shapes defined on the plane. Random sets are characterized by their capacity functionals. Capacity functionals are fuzzy measures. Thus, input images can be integrated with respect to random sets. In this paper, input images are represented as fuzzy sets. The integration is interpreted in the context of mathematical morphology as the average a generalization morphological dilation or erosion. Specifically, the integrals represent the average probability that sets either intersect or are contained in the random sets, the average being over the alpha cuts of the input image. This interpretation has the potential for deriving new learning algorithms for using Choquet integrals in shape detection.

Mathematical morphology has been widely used for shape analysis and feature extraction in digital images. Morphological filters are based on morphological transformations of signals by crisp sets. Their sensitivity to noise led researchers define morphological operations on fuzzy sets. Choquet integrals provide another means of generalizing binary morphological operators. They operate on crisp sets. But, they are referred to as fuzzy integrals because they integrate a real function with respect to a fuzzy measure. An interpretation of a fuzzy measure in terms of fuzzy fitting has led us to generalize gray-scale morphological operators based on Choquet integrals. These operators are called Choquet Morphological Operators (CMO). We used them for domain learning, feature selection, information fusion and, feature extraction. CMOs have been shown to provide better results than conventional morphological operators. In this paper, we give a detailed discussion of the fuzzy fitting concept that leads us to generalizing gray-scale morphological operators using Choquet integrals.

In this paper we develop a hidden Markov model (HMM), called the linguistic HMM (LHMM), suitable for processing sequences of fuzzy vectors. A fuzzy vector B is an n-tuple of fuzzy numbers. Since fuzzy numbers are often associated with linguistic terms, such as "small," "medium," etc., a fuzzy vector can also be called a linguistic vector. The derivation of the linguistic HMM (LHMM) from the numeric HMM is done using the extension principle and the decomposition theorem. We show that the LHMM behaves in the same way as the HMM in the degenerate linguistic case when the fuzzy numbers are singletons (real numbers). We also derive the related algorithms for LHMM training (linguistic Baum-Welch) and for LHMM recognition (linguistic Viterbi). Several examples of LHMM training and recognition are given.

Minimum classification error (MCE) training is proposed to improve performance of a discrete hidden Markov model (DHMM)-based landmine detection system. The system (baseline) was proposed previously for detection of both metal and nonmetal mines from ground-penetrating radar signatures collected by moving vehicles. An initial DHMM model is trained by conventional methods of vector quantization and the Baum-Welch algorithm. A sequential generalized probabilistic descent (GPD) algorithm that minimizes an empirical loss function is then used to estimate the landmine/background DHMM parameters, and an evolutionary algorithm (EA) based on fitness score of classification accuracy is used to generate and select codebooks. The landmine data of one geographical site was used for model training, and those of two different sites were used for evaluation of system performance. Three scenarios were studied: 1) apply MCE/GPD alone to DHMM estimation, 2) apply EA alone to codebook generation, and 3) first apply EA to codebook generation and then apply MCE/GPD to DHMM estimation. Overall, the combined EA and MCE/GPD training led to the best performance. At the same level of detection rate as the baseline DHMM system, the false-alarm rate was reduced by a factor of two, indicating significant performance improvement.

Land mine detection using ground penetrating radar (GPR) is a difficult task because the background clutter characteristics are nonstationary and the land mine signatures are inconsistent. A particularly difficult scenario is the case for which a GPR is mounted on a hand held device with no position or velocity information available to a signal processing algorithm. This paper proposes the use of linear prediction in the frequency domain for land mine detection in this scenario. A frequency domain clutter vector sample is partitioned into subbands. Each subband is modeled by a linear prediction model; the current vector sample is expressed as a linear combination of the past few vector samples plus random noise. The detector first computes the maximum likelihood estimate of the prediction coefficients, and then uses the generalized likelihood method to determine if a land mine is present. The effect of subband processing on the accuracy of the detector is evaluated. Detection results are presented on data collected from a variety of geographical locations. The data sets contain over 2300 mine encounters of different size, shape and content, and a larger number of measurements from locations with no mines. The proposed detector is compared to the baseline differential energy detector. The proposed algorithm reduces the false alarm rate by 60% for all the targets at 90% probability of detection, and 70% for the deep anti-tank mines at 90% probability of detection.

Neural networks (NNs) that are trained to perform classification may not perform as well when used as a module in a larger system. We introduce a novel, system-level method for training NNs with application to counting white blood cells. The idea is to phrase the objective function in terms of total count error rather than the traditional class-coding approach because the goal of this particular recognition system is to accurately count white blood cells of each class, not to classify them. An objective function that represents the sum of the squared counting errors (SSCE) is defined. A batch-mode training scheme based on back-propagation and gradient descent is derived. Sigma and crisp counts are used to evaluate the counting performance. The testing results show that the network trained to minimize SSCE performs better in counting than a classification network with the same structure even though both are trained a comparable number of iterations. This result is consistent with the principle of least commitment of D. Marr (1982).

Traditional hand-written word recognition approaches use character recognition algorithms in the first stage that return confidences in each character under consideration for each segment in the hand-written word. These confidences are used in the second stage to compute a match score between the segmented hand-written word and a given character string. We explore an alternative approach where we view the first stage as a membership assignment process, rather than a character recognition process. The memberships are assigned based on the notion of typicality, and are not relative as in the case of a probabilistic framework. To generate the membership values, we use a robust clustering algorithm that can determine the number of prototypes required to model each character class in a robust and parsimonious manner. Each prototype represents a subclass of the character class. Our experimental results show that when used in conjunction with the proposed approach to word recognition, the memberships generated in this manner produce word recognition results that compare favorably with those of other methods.

Differential white cell counts from bone marrow preparations are very useful in evaluation of various hematologic disorders. It is tedious to locate, identify, and count these classes of cells, even by skilled hands. Automation of classification and counting would be of great benefit. However, the class structure of bone marrow or peripheral blood cells is not discrete; it represents a biological continuum of maturation levels. Because of this, there is uncertainty and overlap in characteristics of adjacent cell classes such that traditional pattern recognition techniques have difficulty in arriving at accurate cell counts. The authors investigate soft counting networks that are trained to produce accurate overall class counts by allowing cells to have degrees of membership in multiple cell classes. This approach is applied to a bone marrow cell library and is compared with other standard recognition algorithms.

As described by Zadeh, recognition technology refers to systems that incorporate new sensors, novel signal processing, and soft computing. In this paper, we discuss these principles applied to the problem of land mine detection. We describe a complex recognition system that is evolving from basic research into a fielded system. Some components of this system have been field tested with excellent results, whereas other components have achieved such results in the laboratory. Fuzzy set-based information fusion algorithms are central to the excellent results obtained. Multiple-detection algorithms are applied to signals acquired from an innovative ground penetrating radar that produces volumetric sub-surface imagery. The outputs of the detection algorithms are combined using the fuzzy logic and Sugeno and Choquet fuzzy integrals to produce overall detection scores. Experimental results are provided on training data and on completely blind test data collected in the field and scored by the US Army.

Novel, general methods for detecting landmine signatures in ground penetrating radar (GPR) using hidden Markov models (HMMs) are proposed and evaluated. The methods are evaluated on real data collected by a GPR mounted on a moving vehicle at three different geographical locations. A large library of digital GPR signatures of both landmines and clutter/background was constructed and used for training. Simple, but effective, observation vector representations are constructed to naturally model the time-varying signatures produced by the interaction of the GPR and the landmines as the vehicle moves. The number and definition of the states of the HMMs are based on qualitative signature models. The model parameters are optimized using the Baum-Welch algorithm. The models were trained on landmine and background/clutter signatures from one geographical location and successfully tested at two different locations. The data used in the test were acquired from over 6000 m/sup 2/ of simulated dirt and gravel roads, and also off-road conditions. These data contained approximately 300 landmine signatures, over half of which were plastic-cased or completely nonmetal.

This experimental study addresses the outlier rejection performance of a multilayer perceptron (MLP) and variations of radial basis function (RBF) networks. Variations include performing principal component decomposition at the RBF centers (PCA-RBF) and adding a regularization term to encourage small variances. Training is performed with and without outliers. The MLPs perform worse than the RBFs. In both cases, the results indicate that if no regularization term is used, then training with outliers can significantly improve the ability of the networks to reject outliers. A significant result is that training the PCA-RBF network with the regularization term and no outlier, we achieve a similar performance as training with outliers.

Presents a method for applying neural networks to the bone marrow white blood cell counting problem. The idea is to phrase the objective function in terms of total count error rather than the traditional class-coding approach. A batch-mode training scheme based on backpropagation and gradient descent is derived. The test results show that, although yielding lower classification rates, the network trained to minimize count error performs better in counting than a classification network with the same structure.

In this paper, we present the theoretical framework for the generalization of classical hidden Markov models using fuzzy measures and fuzzy integrals. The main characteristic of the generalization is the relaxation of the usual additivity constraint of probability measures. Fuzzy integrals are defined with respect to fuzzy measures, whose key property is monotonicity with respect to set inclusion. This property is far weaker than the usual additivity property of probability measures. As a result of the new formulation, the statistical independence assumption of the classical hidden Markov models is relaxed. Two attractive properties of this generalization are: the generalized hidden Markov model reduces to the classical hidden Markov model if we used the Choquet fuzzy integral and probability measures; and the establishment of a relation between the generalized hidden Markov model and the classical nonstationary hidden Markov model in which the transitional parameters vary with time.

For part I see ibid. vol.8, no. 1 (2000). This paper presents an application of the generalized hidden Markov models to handwritten word recognition. The system represents a word image as an ordered list of observation vectors by encoding features computed from each column in the given word image. Word models are formed by concatenating the state chains of the constituent character hidden Markov models. The novel work presented includes the preprocessing, feature extraction, and the application of the generalized hidden Markov models to handwritten word recognition. Methods for training the classical and generalized (fuzzy) models are described. Experiments were performed on a standard data set of handwritten word images obtained from the US Post Office mail stream, which contains real-word samples of different styles and qualities.

Standard shared-weight neural networks previously demonstrated inferior performance to that of morphological shared-weight neural networks for automatic target detection. Empirical analysis showed that entropy measures of the features generated by the standard shared-weight neural networks were consistently lower than those generated by the morphological shared-weight neural networks. Based on this observation, an entropy maximization term was added to the standard shared-weight network objective function. In this paper, we present automatic target detection results for standard shared-weight neural networks trained with and without the added entropy term.

In the standard segmentation-based approach to handwritten word recognition, individual character-class confidence scores are combined via averaging to estimate confidences in the hypothesized identities for a word. We describe a methodology for generating optimal linear combination of order statistics operators for combining character class confidence scores. Experimental results are provided on over 1000 word images.

We remark on an error in the above paper by Chiang (ibid. vol.7 (1999)). The purpose of this note is to present the correct formulas for partial derivatives of fuzzy integral-based neural nodes with respect to densities of Sugeno measures. In reply, Chiang agree with the correction.

Many low-vision people have impaired visual function due to macular scotomas. When scotomas involve the fovea, the low-vision patient must use an eccentric retinal area. Training may help some of these patients compensate for their scotomas. Using a scanning laser ophthalmoscope (SLO), ophthalmologists can measure scotomas and train the patients to use an appropriate retinal area for visual tasks. This task is difficult because the patients may move their eyes consciously or unconsciously during inspection or training. We present a Hough-based registration method combined with feature-based SSD technique to help ophthalmologists perform more accurate and efficient eye registration for inspection and training. Mathematical morphology is used to preprocess each frame to extract a small but relevant set of feature points.

Karyotyping involves the visualization and classification of chromosomes into standard classes. In "normal" human metaphase spreads, chromosomes occur in homologous pairs for the autosomal classes 1-22, and X chromosome for females. Many existing approaches for performing automated human chromosome image analysis presuppose cell normalcy, containing 46 chromosomes within a metaphase spread with two chromosomes per class. This is an acceptable assumption for routine automated chromosome image analysis. However, many genetic abnormalities are directly linked to structural or numerical aberrations of chromosomes within the metaphase spread. Thus, two chromosomes per class cannot be assumed for anomaly analysis. This paper presents the development of image analysis techniques which are extendible to detecting numerical aberrations evolving from structural abnormalities. Specifically, an approach to identifying "normal" chromosomes from selected class(es) within a metaphase spread is presented. Chromosome assignment to a specific class is initially based on neural networks, followed by banding pattern and centromeric index criteria checking, and concluding with homologue matching. Experimental results are presented comparing neural networks as the sole classifier to our homologue matcher for identifying class 17 within normal and abnormal metaphase spreads.

Two hybrid fuzzy neural systems are developed and applied to handwritten word recognition. The word recognition system requires a module that assigns character class membership values to segments of images of handwritten words. The module must accurately represent ambiguities between character classes and assign low membership values to a wide variety of noncharacter segments resulting from erroneous segmentations. Each hybrid is a cascaded system. The first stage of both is a self-organizing feature map (SOFM). The second stages map distances into membership values. The third stage of one system is a multilayer perceptron (MLP). The third stage of the other is a bank of Choquet fuzzy integrals (FI). The two systems are compared individually and as a combination to the baseline system. The new systems each perform better than the baseline system. The MLP system slightly outperforms the FI system, but the combination of the two outperforms the individual systems with a small increase in computational cost over the MLP system. Recognition rates of over 92% are achieved with a lexicon set having average size of 100. Experiments were performed on a standard test set from the SUNY/USPS CD-ROM database.

A shared-weight neural network based on mathematical morphology is introduced. The feature extraction process is learned by interaction with the classification process. Feature extraction is performed using gray-scale hit-miss transforms that are independent of gray-level shifts. The morphological shared-weight neural network (MSNN) is applied to automatic target recognition. Two sets of images of outdoor scenes are considered. The first set consists of two subsets of infrared images of tracked vehicles. The goal in this set is to reject the background and to detect tracked vehicles. The second set consists of visible images of cars in a parking lot. The goal in this set is to detect the Chevrolet Blazers with various degrees of occlusion. A training method that is effective in reducing false alarms and a target aim point selection algorithm are introduced. The MSNN is compared to the standard shared-weight neural network. The MSNN trains relatively quickly and exhibits better generalization.

An off-line handwritten word recognition system is described. Images of handwritten words are matched to lexicons of candidate strings. A word image is segmented into primitives. The best match between sequences of unions of primitives and a lexicon string is found using dynamic programming. Neural networks assign match scores between characters and segments. Two particularly unique features are that neural networks assign confidence that pairs of segments are compatible with character confidence assignments and that this confidence is integrated into the dynamic programming. Experimental results are provided on data from the U.S. Postal Service.

A hybrid feature extraction framework for handwritten numeric fields recognition is described. The numeric fields were extracted from binary images of credit card application forms. The images include identity numbers (ID) and phone numbers. The feature extraction framework utilizes a cascade of a Kohonen self-organizing feature map (SOM) and a set of non-linear filtering units. The goals of our feature extraction process are to provide reliable information to the recognition stage. The recognition stage uses the feature set as inputs to a multi-layer neural network. We present experimental results which demonstrate the ability to extract features automatically in handwritten digit recognition. Experiments were performed on a test data set from the CCL/ITRI Database which consists of over 90,390 handwritten numeric digits. Recognition rate of 98.6% is achieved on this database.

An automatic feature generation method for handwritten digit recognition is described. Two different evaluation measures, orthogonality and information, are used to guide the search for features. The features are used in a backpropagation trained neural network. Classification rates compare favorably with results published in a survey of high-performance handwritten digit recognition systems. This classifier is combined with several other high performance classifiers. Recognition rates of around 98% are obtained using two classifiers on a test set with 1000 digits per class.

A lexicon-based, handwritten word recognition system combining segmentation-free and segmentation-based techniques is described. The segmentation-free technique constructs a continuous density hidden Markov model for each lexicon string. The segmentation-based technique uses dynamic programming to match word images and strings. The combination module uses differences in classifier capabilities to achieve significantly better performance.

One of the longest standing problems in medical image analysis is that of the automated recognition of chromosomes from images of a metaphase spread of a cell. This process of visualizing and categorizing the chromosomes within a cell, called karyotyping, is a key factor in many medical procedures. It is a labor intensive activity, and hence, is a great candidate for automation. However, there are many sources of uncertainty in this problem domain, making complete karyotyping a difficult problem. We describe how fuzzy logic is being inserted into a complete karyotyping system to deal with uncertainty in similar chromosome classes.>

Fuzzy logic is applied to the problem of locating and reading street numbers in digital images of handwritten mail. A fuzzy rule-based system is defined that uses uncertain information provided by image processing and neural network-based character recognition modules to generate multiple hypotheses with associated confidence values for the location of the street number in an image of a handwritten address. The results of a blind test of the resultant system are presented to demonstrate the value of this new approach. The results are compared to those obtained using a neural network trained with backpropagation. The fuzzy logic system achieved higher performance rates.>

Experiments comparing neural networks trained with crisp and fuzzy desired outputs are described. A handwritten word recognition algorithm using the neural networks for character level confidence assignment was tested on images of words taken from the United States Postal Service mailstream. The fuzzy outputs were defined using a fuzzy k-nearest neighbor algorithm. The crisp networks slightly outperformed the fuzzy networks at the character level but the fuzzy networks outperformed the crisp networks at the word level. This empirical result is interpreted as an example of the principle of least commitment.>