The current standard for training brain-computer interface (BCI) machine learning models is user-specific. There is a high interest in developing generic models that are trained on data from other users to minimize BCI calibration time; however, this is limited by noisy, non-stationary brain signals and high inter-user variability. We investigate the trade-off between training data quality and quantity on P300 BCI performance in individuals with amyotrophic lateral sclerosis (ALS) with representative traditional machine learning (stepwise linear discriminant analysis, SWLDA) and deep learning (EEGNet) models. Results show that data quality and domain alignment are more critical than dataset size: user-specific models trained on significantly less data outperformed generic models; generic models trained on ALS data outperformed models trained on non-ALS data; block-averaging of features was mostly detrimental to EEGNet but beneficial to SWLDA; and accounting for inter-stimulus interval differences between ALS and non-ALS data had minimal effect. Our findings highlight the importance of individualized model tuning for reliable P300 BCIs.

Brain-computer interfaces (BCIs) can provide a means of communication for individuals with severe neuromuscular diseases, the target end-users. While personalized BCI machine learning models are the current standard, models trained on data from other users could reduce BCI calibration time. We use a novel dataset with BCI users with and without amyotrophic lateral sclerosis (ALS) and a popular BCI deep learning model, EEGNet, to assess the impact of population domain data on transfer learning of a P300 speller task in the ALS cohort. Results show that training on source data from the non-ALS cohort was detrimental to transfer learning. In contrast, generic EEGNet models trained on source data from the ALS cohort performed comparably as user-specific models. Our findings highlight the need for more data from target end-users populations in publicly available BCI datasets.

Mixture-of-Experts (MoE) models have become popular in machine learning, boosting performance by partitioning tasks across multiple experts. However, the need for several experts often results in high computational costs, limiting their application on resource-constrained devices with stringent real-time requirements, such as cochlear implants (CIs). We introduce the Omni-Expert (OE) – a simple and efficient solution that leverages feature transformations to achieve the 'divide-and-conquer' functionality of a full MoE ensemble in a single expert model. We demonstrate the effectiveness of the OE using phoneme-specific time-frequency masking for speech dereverberation in a CI. Empirical results show that the OE delivers statistically significant improvements in objective intelligibility measures of CI vocoded speech at different levels of reverberation across various speech datasets at a much reduced computational cost relative to a counterpart MoE.

Recent object detection models for infrared (IR) imagery are based upon deep neural networks (DNNs) and require large amounts of labeled training imagery. However, publicly available datasets that can be used for such training are limited in their size and diversity. To address this problem, we explore cross-modal style transfer (CMST) to leverage large and diverse color imagery datasets so that they can be used to train DNN-based IR image-based object detectors. We evaluate six contemporary stylization methods on four publicly-available IR datasets - the first comparison of its kind - and find that CMST is highly effective for DNN-based detectors. Surprisingly, we find that existing data-driven methods are outperformed by a simple grayscale stylization (an average of the color channels). Our analysis reveals that existing data-driven methods are either too simplistic or introduce significant artifacts into the imagery. To overcome these limitations, we propose meta-learning style transfer (MLST), which learns a stylization by composing and tuning well-behaved analytic functions. We find that MLST leads to more complex stylizations without introducing significant image artifacts and achieves the best overall detector performance on our benchmark datasets.

Recently, the first foundation model developed specifically for image segmentation tasks was developed, termed the "Segment Anything Model" (SAM). SAM can segment objects in input imagery based on cheap input prompts, such as one (or more) points, a bounding box, or a mask. The authors examined the zero-shot image segmentation accuracy of SAM on a large number of vision benchmark tasks and found that SAM usually achieved recognition accuracy similar to, or sometimes exceeding, vision models that had been trained on the target tasks. The impressive generalization of SAM for segmentation has major implications for vision researchers working on natural imagery. In this work, we examine whether SAM’s performance extends to overhead imagery problems and help guide the community’s response to its development. We examine SAM’s performance on a set of diverse and widely studied benchmark tasks. We find that SAM does often generalize well to overhead imagery, although it fails in some cases due to the unique characteristics of overhead imagery and its common target objects. We report on these unique systematic failure cases for remote sensing imagery that may comprise useful future research for the community.

We propose and show the efficacy of a new method to address generic inverse problems. Inverse modeling is the task whereby one seeks to determine the hidden parameters of a natural system that produce a given set of observed measurements. Recent work has shown impressive results using deep learning, but we note that there is a trade-off between model performance and computational time. For some applications, the computational time at inference for the best performing inverse modeling method may be overly prohibitive to its use. In seeking a faster, high-performing model, we present a new method that leverages multiple manifolds as a mixture of backward (e.g., inverse) models in a forward-backward model architecture. These multiple backwards models all share a common forward model, and their training is mitigated by generating training examples from the forward model. The proposed method thus has two innovations: 1) the multiple Manifold Mixture Network (MMN) architecture, and 2) the training procedure involving augmenting backward model training data using the forward model. We demonstrate the advantages of our method by comparing to several baselines on four benchmark inverse problems, and we furthermore provide analysis to motivate its design.

There is evidence that transformers offer state-of-the-art recognition performance on tasks involving overhead imagery (e.g., satellite imagery). However, it is difficult to make unbiased empirical comparisons between competing deep learning models, making it unclear whether, and to what extent, transformer-based models are beneficial. In this paper we systematically compare the impact of adding transformer structures into state-of-the-art segmentation models for overhead imagery. Each model is given a similar budget of free parameters, and their hyperparameters are optimized using Bayesian Optimization with a fixed quantity of data and computation time. We conduct our experiments with a large and diverse dataset comprising two large public benchmarks: Inria and DeepGlobe. We perform additional ablation studies to explore the impact of specific transformer-based modeling choices. Our results suggest that transformers provide consistent, but modest, performance improvements. We only observe this advantage however in hybrid models that combine convolutional and transformer-based structures, while fully transformer-based models achieve relatively poor performance.

Brain-computer interfaces (BCIs), such as the P300 speller, can provide a means of communication for individuals with severe neuromuscular limitations. BCIs interpret electroencephalography (EEG) signals in order to translate embedded information about a user's intent into executable commands to control external devices. However, EEG signals are inherently noisy and nonstationary, posing a challenge to extended BCI use. Conventionally, a BCI classifier is trained via supervised learning in an offline calibration session; once trained, the classifier is deployed for online use and is not updated. As the statistics of a user's EEG data change over time, the performance of a static classifier may decline with extended use. It is therefore desirable to automatically adapt the classifier to current data statistics without requiring offline recalibration. In an existing semi-supervised learning approach, the classifier is trained on labeled EEG data and is then updated using incoming unlabeled EEG data and classifier-predicted labels. To reduce the risk of learning from incorrect predictions, a threshold is imposed to exclude unlabeled data with low-confidence label predictions from the expanded training set when retraining the adaptive classifier. In this work, we propose the use of a language model for spelling error correction and disambiguation to provide information about label correctness during semi-supervised learning. Results from simulations with multi-session P300 speller user EEG data demonstrate that our language-guided semi-supervised approach significantly improves spelling accuracy relative to conventional BCI calibration and threshold-based semi-supervised learning.

Speech intelligibility in cochlear implant (CI) users degrades considerably in listening environments with reverberation and noise. Previous research in automatic speech recognition (ASR) has shown that phoneme-based speech enhancement algorithms improve ASR system performance in reverberant environments as compared to a global model. However, phoneme-specific speech processing has not yet been implemented in CIs. In this paper, we propose a causal deep learning framework for classifying phonemes using features extracted at the time-frequency resolution of a CI processor. We trained and tested long short-term memory networks to classify phonemes and manner of articulation in anechoic and reverberant conditions. The results showed that CI-inspired features provide slightly higher levels of performance than traditional ASR features. To the best of our knowledge, this study is the first to provide a classification framework with the potential to categorize phonetic units in real-time in a CI.

Thermal infrared (IR) imaging has increasingly been used for remote sensing applications, which has required the adaptation of processing techniques for “natural images” (e.g., RGB images) to this unique domain in order to accommodate such differences as texture and resolution. While Convolutional Neural Networks (CNNs) have recently been shown to perform well for classification of IR images, we consider the common scenario in which the target object is partially occluded. Recent work demonstrates that deep CNNs struggle to generalize under occlusion, and Compositional CNNs (CompNets) have been proposed as deep models to mitigate this shortcoming. In this work, we apply CompNets to IR imagery, discuss the considerations in moving from natural images to IR, and analyze performance on a dataset that has been artificially occluded. Our results indicate that CompNets do, indeed, bolster robustness to occlusion in the IR domain.

Cochlear implant (CI) users experience substantial difficulties in understanding reverberant speech. A previous study proposed a strategy that leverages automatic speech recognition (ASR) to recognize reverberant speech and speech synthesis to translate the recognized text into anechoic speech. However, the strategy was trained and tested on the same reverberant environment, so it is unknown whether the strategy is robust to unseen environments. Thus, the current study investigated the performance of the previously proposed algorithm in multiple unseen environments. First, an ASR system was trained on anechoic and reverberant speech using different room types. Next, a speech synthesizer was trained to generate speech from the text predicted by the ASR system. Experiments were conducted in normal hearing listeners using vocoded speech, and the results showed that the strategy improved speech intelligibility in previously unseen conditions. These results suggest that the ASR-synthesis strategy can potentially benefit CI users in everyday reverberant environments.

Access to electricity positively correlates with many beneficial socioeconomic outcomes in the developing world including improvements in education, health, and poverty. Efficient planning for electricity access requires information on the location of existing electric transmission and distribution infrastructure; however, the data on existing infrastructure is often unavailable or expensive. We propose a deep learning based method to automatically detect electric transmission infrastructure from aerial imagery and quantify those results with traditional object detection performance metrics. In addition, we explore two challenges to applying these techniques at scale: (1) how models trained on particular geographies generalize to other locations and (2) how the spatial resolution of imagery impacts infrastructure detection accuracy. Our approach results in object detection performance with an F1 score of 0.53 (0.47 precision and 0.60 recall). Using training data that includes more diverse geographies improves performance across the 4 geographies that we examined. Image resolution significantly impacts object detection performance and decreases precipitously as the image resolution decreases.

Recently, Convolutional Neural Networks (CNNs) have demonstrated impressive performance on several visual recognition benchmark datasets utilizing overhead imagery. However, most of these analyses performed on benchmark datasets involve testing pre-trained CNNs on imagery that was collected over roughly the same locations as the training imagery. In this work we propose a benchmark problem - termed cross-geographical domain (xGD) adaptation - designed to evaluate the performance of CNNs in which they are tested on imagery collected over previously unseen geo-locations - a more challenging and practical scenario that we term cross-domain testing. We focus this work on building segmentation due to the availability of appropriate datasets. The results indicate that CNNs generalize poorly to data processed from geographic locations that were not present in training. Surprisingly, we found that larger models (pre-trained on ImageNet) generalize as well as small models in cross-domain testing, and sometimes better. This work provides the first comprehensive results for cross-domain recognition, raising awareness of this important problem. We hope that xGD can serve as a benchmark for future work; xGD uses publicly-available data, and we release our design details with this publication.

Segmentation convolutional neural networks (CNNs) are now popular for the semantic segmentation (i.e., dense pixel-wise labeling) of remote sensing imagery, such as color or hyperspectral satellite imagery. In recent years a large number of hand-labeled datasets of overhead imagery have emerged, leading to breakthrough performance for CNNs. However, these datasets are typically used in isolation of one another because they are either (i) annotated with heterogeneous object type labels, or (ii) they are collected over different geographic areas. This imposes a major bottleneck on the value of these datasets. In this work we present what we call a class-asymmetric loss function that makes it possible to train a single multi-class network using multiple datasets that are heterogeneously-labeled. We show, for example, that it is possible to train a segmentation algorithm for Buildings, roads, and background using two datasets: one annotated with buildings and one annotated with buildings. We propose a class asymmetric loss that under certain common conditions, allows for one to train models on datasets in which the target class is unlabeled.

Segmentation convolutional neural networks (SCNNs) are now popular for the semantic segmentation (i.e., dense pixel-wise labeling) of remote sensing imagery, such as color or hyperspectral satellite imagery. One desirable property of SCNNs when applied to remote sensing problems is rotational equivariance. This property implies that the class label assigned to a particular pixel (building, road, etc.) does not change if the input imagery is rotated by an arbitrary angle. We argue that recently proposed methods to make rotational equivariant SCNNs fall into two broad categories: easily employed methods that are somewhat ineffective, and highly effective methods that are complicated and potentially incompatible with state-of-the-art SCNN techniques. We propose a simple addition to the standard SCNN loss function that encourages the SCNN to be rotationally equivariant, and is easily added to modern SCNNs. We test the method on the Inria building labeling dataset and compare it to the popular simple approach of adding random rotational augmentations of the input imagery during training. We show that the proposed approach (i) achieves improved equivariance and (ii) yields performance improvements on average.

In this paper, we consider the development of algorithms for the automatic detection of buried threats using ground penetrating radar (GPR) measurements. GPR is one of the most studied and successful modalities for automatic buried threat detection (BTD), and a large variety of BTD algorithms have been proposed for it. Despite this, large-scale comparisons of GPR-based BTD algorithms are rare in the literature. In this paper, we report the results of a multi-institutional effort to develop advanced BTD algorithms for a real-world GPR BTD system. The effort involved five institutions with substantial experience with the development of GPR-based BTD algorithms. In this paper, we report the technical details of the advanced algorithms submitted by each institution, representing their latest technical advances, and many state-of-the-art GPR-based BTD algorithms. We also report the results of evaluating the algorithms from each institution on the large experimental data set used for development. The experimental data set comprised 120 000 m2of GPR data using surface area, from 13 different lanes across two U.S. test sites. The data were collected using a vehicle-mounted GPR system, the variants of which have supplied data for numerous publications. Using these results, we identify the most successful and common processing strategies among the submitted algorithms, and make recommendations for GPR-based BTD algorithm design.

Latent Dirichlet allocation (LDA) is a statistical model that is often used to discover topics or themes in a large collection of documents. In the LDA model, topics are modeled as discrete distributions over a finite vocabulary of words. The LDA is also a popular choice to model other datasets spanning a discrete domain, such as population genetics and social networks. However, in order to model data spanning a continuous domain with the LDA, discrete approximations of the data need to be made. These discrete approximations to continuous data can lead to loss of information and may not represent the true structure of the underlying data. We present an augmented version of the LDA topic model, where topics are represented using Gaussian mixture models (GMMs), which are multi-modal distributions spanning a continuous domain. This augmentation of the LDA topic model with Gaussian mixture topics is denoted by the GMM-LDA model. We use Gibbs sampling to infer model parameters. We demonstrate the utility of the GMM-LDA model by applying it to the problem of clustering sleep states in electroencephalography (EEG) data. Results are presented demonstrating superior clustering performance with our GMM-LDA algorithm compared to the standard LDA and other clustering algorithms.

Individuals with cochlear implants (CIs) experience more difficulty understanding speech in reverberant environments than normal hearing listeners. As a result, recent research has targeted mitigating the effects of late reverberant signal reflections in CIs by using a machine learning approach to detect and delete affected segments in the CI stimulus pattern. Previous work has trained electrode-specific classification models to mitigate late reverberant signal reflections based on features extracted from only the acoustic activity within the electrode of interest. Since adjacent CI electrodes tend to be activated concurrently during speech, we hypothesized that incorporating additional information from the other electrode channels, termed cross-channel information, as features could improve classification performance. Cross-channel information extracted in real-world conditions will likely contain errors that will impact classification performance. To simulate extracting cross-channel information in realistic conditions, we developed a graphical model based on the Ising model to systematically introduce errors to specific types of cross-channel information. The Ising-like model allows us to add errors while maintaining the important geometric information contained in cross-channel information, which is due to the spectro-temporal structure of speech. Results suggest the potential utility of leveraging cross-channel information to improve the performance of the reverberation mitigation algorithm from the baseline channel-based features, even when the cross-channel information contains errors.

In this work, we investigate strategies for training convolutional neural networks (CNNs) to perform recognition on remote sensing imagery. In particular we consider the particular problem of semantic segmentation in which the goal is to obtain a dense pixel-wise labeling of the input imagery. Remote sensing imagery is usually stored in the form of very large images, called “tiles”, which are too big to be segmented directly using most CNNs and their associated hardware. Therefore smaller sub-images, called “patches”, must be extracted from the available tiles. A popular strategy in the literature is to randomly sample patches from the tiles. However, in this work we demonstrate experimentally that extracting patches randomly from a uniform, non-overlapping spatial grid, leads to more accurate models. Our findings suggest the performance improvements are the result of reducing redundancy within the training dataset. We also find that sampling mini-batches of patches (for stochastic gradient descent) using constraints that maximizes the diversity of images within each batch leads to more accurate models. For example, in this work we constrained patches to come from varying tiles, or cities. These simple strategies contributed to our winning entry (in terms of overall performance) in the first year of the INRIA Building Labeling Challenge.

In this work we consider the application of convolutional neural networks (CNNs) for the semantic segmentation of remote sensing imagery (e.g., aerial color or hyperspectral imagery). In segmentation the goal is to provide a dense pixel-wise labeling of the input imagery. However, remote sensing imagery is usually stored in the form of very large images, called “tiles”, which are too large to be segmented directly using most CNNs and their associated hardware. During label inference (i.e., obtaining labels for a new large tile) smaller sub-images, called “patches”, are extracted uniformly over a tile and the resulting label maps are “stitched” (or concatenated) to create a tile-sized label map. This approach suffers from computational inefficiency and risks of discontinuities at the boundaries between the output of individual patches. In this work we propose a simple alternative approach in which the input size of the CNN is dramatically increased only during label inference. We evaluate the performance of the proposed approach against a standard stitching approach using two popular segmentation CNN models on the INRIA building labeling dataset. The results suggest that the proposed approach substantially reduces label inference time, while also yielding modest overall label accuracy increases. This approach also contributed to our winning entry (overall performance) in the INRIA building labeling competition.

Over the recent years, there has been an increasing interest in large-scale classification of remote sensing images. In this context, the Inria Aerial Image Labeling Benchmark has been released online in December 2016. In this paper, we discuss the outcomes of the first year of the benchmark contest, which consisted in dense labeling of aerial images into building / not building classes, covering areas of five cities not present in the training set. We present four methods with the highest numerical accuracies, all four being convolutional neural network approaches. It is remarkable that three of these methods use the U-net architecture, which has thus proven to become a new standard in image dense labeling.

Recently, convolutional neural networks (CNNs) have received substantial attention in the literature for object recognition (e.g., buildings and roads) in several remote sensing data modalities (e.g., aerial color imagery). Although CNNs have exhibited excellent recognition performance, recent research suggests that trained CNNs can often perform very poorly when applied to data collected over new geographic regions, and for which little labeled training data is available. In this work, we consider the adversarial discriminative domain adaptation (ADDA) approach to address this limitation, due its recent success on related problems. A limitation of ADDA is that it is unsupervised, so in this work we extend ADDA to a semi-supervised algorithm, in which we assume that both labeled and unlabeled data are available in the new domain (e.g., in new geographic region to be evaluated). We compare semi-supervised ADDA to ADDA and a standard fine-tuning approach wherein available labeled data is used for standard CNN training. We perform experiments on two remote sensing datasets and the results indicate that semi-supervised ADDA consistently improves over the other approaches when small amounts of labeled training data are available in the new domain.

Stimulus-driven brain-computer interfaces (BCIs), such as the P300 speller, rely on using a sequence of sensory stimuli to elicit specific neural responses as control signals, while a user attends to relevant target stimuli that occur within the sequence. In current BCIs, the stimulus presentation schedule is typically generated in a pseudo-random fashion. Given the non-stationarity of brain electrical signals, a better strategy could be to adapt the stimulus presentation schedule in real-time by selecting the optimal stimuli that will maximize the signal-to-noise ratios of the elicited neural responses and provide the most information about the user's intent based on the uncertainties of the data being measured. However, the high-dimensional stimulus space limits the development of algorithms with tractable solutions for optimized stimulus selection to allow for real-time decision-making within the stringent time requirements of BCI processing. We derive a simple analytical solution of an information-based objective function for BCI stimulus selection by transforming the high-dimensional stimulus space into a one-dimensional space that parameterizes the objective function - the prior probability mass of the stimulus under consideration, irrespective of its contents. We demonstrate the utility of our adaptive stimulus selection algorithm in improving BCI performance with results from simulation and real-time human experiments.

Immersive virtual reality (VR) systems offer flexible control of an interactive environment, along with precise position and orientation tracking of realistic movements. Immersive VR can also be used in conjunction with neurophysiological monitoring techniques, such as electroencephalography (EEG), to record neural activity as users perform complex tasks. As such, the fusion of VR, kinematic tracking, and EEG offers a powerful testbed for naturalistic neuroscience research. In this study, we combine these elements to investigate the cognitive and neural mechanisms that underlie motor skill learning during a multi-day simulated marksmanship training regimen conducted with 20 participants. On each of 3 days, participants performed 8 blocks of 60 trials in which a simulated clay pigeon was launched from behind a trap house. Participants attempted to shoot the moving target with a firearm game controller, receiving immediate positional feedback and running scores after each shot. Over the course of the 3 days that individuals practiced this protocol, shot accuracy and precision improved significantly while reaction times got significantly faster. Furthermore, results demonstrate that more negative EEG amplitudes produced over the visual cortices correlate with better shooting performance measured by accuracy, reaction times, and response times, indicating that early visual system plasticity underlies behavioral learning in this task. These findings point towards a naturalistic neuroscience approach that can be used to identify neural markers of marksmanship performance.

Forward-looking ground-penetrating radar (FLGPR) has recently been investigated as a remote-sensing modality for buried target detection (e.g., landmines). In this context, raw FLGPR data are beamformed into images, and then, computerized algorithms are applied to automatically detect subsurface buried targets. Most existing algorithms are supervised, meaning that they are trained to discriminate between labeled target and nontarget imagery, usually based on features extracted from the imagery. A large number of features have been proposed for this purpose; however, thus far it is unclear as to which are the most effective. The first goal of this paper is to provide a comprehensive comparison of detection performance using existing features on a large collection of FLGPR data. Fusion of the decisions resulting from processing each feature is also considered. The second goal of this paper is to investigate two modern feature learning approaches from the object recognition literature: the bag-of-visual words and the Fisher vector for FLGPR processing. The results indicate that the new feature learning approaches lead to the best performing FLGPR algorithm. The results also show that fusion between existing features and new features yields no additional performance improvements.

Ground-penetrating radar (GPR) is one of the most popular and successful sensing modalities that have been investigated for landmine and subsurface threat detection. Many of the detection algorithms applied to this task are supervised and therefore require labeled examples of threat and nonthreat data for training. Training data most often consist of 2-D images (or patches) of GPR data, from which features are extracted and provided to the classifier during training and testing. Identifying desirable training and testing locations to extract patches, which we term “keypoints,” is well established in the literature. In contrast, however, a large variety of strategies have been proposed regarding keypoint utilization (e.g., how many of the identified keypoints should be used at threat, or nonthreat, locations). Given a variety of keypoint utilization strategies that are available, it is very unclear: 1) which strategies are best or 2) whether the choice of strategy has a large impact on classifier performance. We address these questions by presenting a taxonomy of existing utilization strategies and then evaluating their effectiveness on a large data set using many different classifiers and features. We analyze the results and propose a new strategy, called PatchSelect, which outperforms other strategies across all experiments.

The P300-based brain-computer interface (BCI) speller relies on eliciting and detecting specific brain responses to target stimulus events, termed event-related potentials (ERPs). In a visual speller, ERPs are elicited when the user's desired character, i.e. the “target,” is flashed on a computer screen. The P300 speller is currently limited by its relatively slow typing speed due to the need for repetitive data measurements that are necessary to achieve reasonable signal-to-noise ratios. In addition, refractory effects limit the ability to elicit ERPs with every target stimulus event presentation. In this paper, we present a new method to design the stimulus presentation paradigm for the P300 speller by exploiting an information-theoretic approach to maximize the information content that is presented to the user while also mitigating refractory effects. We present results with real-time BCI use which demonstrate significant performance improvements with our performance-based paradigm compared to the conventional stimulus presentation paradigm.

In this work we consider the problem of developing algorithms that automatically identify small-scale solar photovoltaic arrays in high resolution aerial imagery. Such algorithms potentially offer a faster and cheaper solution to collecting small-scale photovoltaic (PV) information, such as their location, capacity, and the energy they produce. Here we build on previous algorithmic work by employing convolutional neural networks (CNNs), which have recently yielded major improvements in other image object recognition problems. We propose a CNN architecture for our recognition problem and then measure its detection performance on the same (publicly available) dataset that was used in previous publications. The results indicate that the CNN yields substantial performance improvements over previous results. We also investigate the recently popular approach of pre-training for CNNs.

In this work, we consider the problem of detecting target objects in remote sensing imagery; such as detecting rooftops, trees, or cars in color/hyperspectral imagery. Many detection algorithms for this problem work by assigning a decision statistic (or “confidence”) to all, or a subset, of spatial locations in the data. A threshold is then applied to the statistics to identify detections. The detection theory underpinning this general approach assumes that a yes/no decision must be made, individually, for each location. In some applications, however, the precise location of the detected objects may be less important than knowing how many total objects there are. In this work we propose two methods that can permit a generic detection algorithm to gradually lower the spatial certainty, or resolution, of its detections in order to improve the accuracy of the overall number of detected objects. We validate the proposed methods on a controlled synthetic dataset as well as a real dataset from previously published work on solar photovoltaic array detection in color aerial imagery.

In this work, the problem of developing algorithms that automatically infer information about small-scale solar photovoltaic (PV) arrays in high resolution aerial imagery is considered. Such algorithms potentially offer a faster and cheaper solution to collecting small-scale PV information, such as their location and capacity. Existing work on this topic has focused on the automatic identification and annotation of panels in the aerial imagery. We extend this work by showing that we can reliably infer the capacity of PV arrays given only (i) color aerial imagery and (ii) a precise annotation of the array location. First we demonstrate that accurate capacity estimates can be obtained simply by estimating the visible surface area of a solar array, regardless of tilt. We then build a more sophisticated model where we use additional image information related to the properties of the solar array to further improve the capacity predictions. We use a dataset of 362 manually annotated Google Earth images of solar arrays with known electricity generation capacity in North Carolina to measure the predictive performance of our models.

Brain-computer interfaces (BCIs) can provide an alternative means of communication for individuals with severe neuromuscular limitations. The P300-based BCI speller relies on eliciting and detecting transient event-related potentials (ERPs) in electroencephalography (EEG) data, in response to a user attending to rarely occurring target stimuli amongst a series of non-target stimuli. However, in most P300 speller implementations, the stimuli to be presented are randomly selected from a limited set of options and stimulus selection and presentation are not optimized based on previous user data. In this work, we propose a data-driven method for stimulus selection based on the expected discrimination gain metric. The data-driven approach selects stimuli based on previously observed stimulus responses, with the aim of choosing a set of stimuli that will provide the most information about the user's intended target character. Our approach incorporates knowledge of physiological and system constraints imposed due to real-time BCI implementation. Simulations were performed to compare our stimulus selection approach to the row-column paradigm, the conventional stimulus selection method for P300 spellers. Results from the simulations demonstrated that our adaptive stimulus selection approach has the potential to significantly improve performance from the conventional method: up to 34% improvement in accuracy and 43% reduction in the mean number of stimulus presentations required to spell a character in a 72-character grid. In addition, our greedy approach to stimulus selection provides the flexibility to accommodate design constraints.

The automatic detection of buried threats in ground penetrating radar (GPR) data is an active area of research due to GPR's ability to detect both metal and nonmetal subsurface objects. Recent work on algorithms designed to distinguish between threats and nonthreats in GPR data has utilized computer vision methods to advance the state-of-the-art detection and discrimination performance. Feature extractors, or descriptors, from the computer vision literature have exhibited excellent performance in representing 2-D GPR image patches and allow for robust classification of threats from nonthreats. This paper aims to perform a broad study of feature extraction methods in order to identify characteristics that lead to improved classification performance under controlled conditions. The results presented in this paper show that gradient-based features, such as the edge histogram descriptor and the scale invariant feature transform, provide the most robust performance across a large and varied data set. These results indicate that various techniques from the computer vision literature can be successfully applied to target detection in GPR data and that more advanced techniques from the computer vision literature may provide further performance improvements.

Most state-of-the-art dictionary learning algorithms (DLAs) are iterative, and must begin with an initial estimate of the dictionary, referred to as the seed. A seed can be generated randomly, but it has been shown that choosing a more intelligent seed often yields a better solution. For example, a seed inferred using data from a related problem, or one handcrafted based on a priori knowledge of the problem at hand can yield better solutions. Seed dictionaries appear to encode valuable a priori information however, most DLAs discard the seed after initialization. This work investigates the questions of whether the information encoded in a good seed can be leveraged further, by potentially using the seed to influence learning after initialization. This is achieved by modifying the popular DLA K-SVD to use the seed as a prior during learning, by penalizing differences between the learned dictionary and the seed. The resulting algorithm, referred to as Seed Shrinkage Dictionary Learning (SSDL), is examined against K-SVD on image denoising experiments using several benchmark images. The results indicate that utilizing the seed as a prior in this way consistently yields improved denoising performance in our experiments. This simple approach motivates the development of more sophisticated approaches that leverage a priori information in useful seeds.

The ground penetrating radar (GPR) is a popular and successful remote sensing modality that has been investigated for landmine detection. GPR offers excellent detection performance, but it is limited by a low rate of advance (ROA) due to its short sensing standoff distance. Standoff distance refers to the distance between the sensing platform and the location in front of the platform where the GPR senses the ground. Large standoff (high ROA) sensing modalities have been investigated as alternatives to the GPR, but they do not yet achieve comparable detection performance. This paper proposes a new sensor management approach, called multistate management (MSM), which combines large and short standoff sensors on the same platform in a way that leverages their respective advantages, yielding a system with better ROA and detection performance. MSM is more difficult to analyze than traditional systems because it allows sensor activity and system velocity to change over time. Therefore, a new probabilistic model based on queuing theory, called Q-MSM, is also proposed for analyzing and designing detection systems operating with MSM. Simulations were conducted using real field-collected data for a system with a large standoff forward-looking infrared camera and a GPR. The system is operated with MSM, and the results show that this leads to better ROA and detection performance than can be attained otherwise. Furthermore, the results show that Q-MSM can accurately predict the behavior of the MSM system, validating its utility for analyzing and designing such systems.

Multiple instance learning (MIL) is a type of supervised learning in which labels are available for sets of observations (bags), but not for individual observations (instances). MIL has been applied in different areas, which has led to a large number of algorithms for learning based on MIL data. Many of these approaches focus on maximizing class margins, performing instance selection, or developing distance metrics and kernels suitable for application directly to bags. Although these approaches have shown promise, most require cross-validation-based optimization of hyper parameters or iterative numerical optimization to determine the proper number of target concepts. This work proposes a nonparametric Bayesian approach to learning in MIL scenarios based on Dirichlet process mixture models. The nonparametric nature of the model and the use of noninformative priors remove the need to perform cross-validation-based optimization while variational Bayesian inference allows for rapid parameter learning. The resulting approach generalizes to different applications by easily incorporating alternate data generation models. In a related effort [A. Manandhar et al., IEEE Trans. Geosci. Remote Sensing53(4) (2015) 1737–1745.], the proposed model has been extended to incorporate time-varying data. Results indicate that when the data generation assumption holds, the proposed approach performs competitively with existing MIL and nonMIL methods for several standard MIL datasets and a new MIL dataset introduced in this work.

Hidden Markov models (HMMs) have previously been successfully applied to subsurface threat detection using ground penetrating radar (GPR) data. However, parameter estimation in most HMM-based landmine detection approaches is difficult since object locations are typically well known for the 2-D coordinates on the Earth's surface but are not well known for object depths underneath the ground/time of arrival in a GPR A-scan. As a result, in a standard expectation maximization HMM (EM-HMM), all depths corresponding to a particular alarm location may be labeled as target sequences although the characteristics of data from different depths are substantially different. In this paper, an alternate HMM approach is developed using a multiple-instance learning (MIL) framework that considers an unordered set of HMM sequences at a particular alarm location, where the set of sequences is defined as positive if at least one of the sequences is a target sequence; otherwise, the set is defined as negative. Using the MIL framework, a collection of these sets (bags), along with their labels is used to train the target and nontarget HMMs simultaneously. The model parameters are inferred using variational Bayes, making the model tractable and computationally efficient. Experimental results on two synthetic and two landmine data sets show that the proposed approach performs better than a standard EM-HMM.

Many remote sensing applications involve the classification of anomalous responses as either objects of interest or clutter. This paper addresses the problem of anomaly classification in hyperspectral imagery (HSI) and focuses on robustly detecting disturbed earth in the long-wave infrared (LWIR) spectrum. Although disturbed earth yields a distinct LWIR signature that distinguishes it from the background, its distribution relative to clutter may vary over different environmental contexts. In this paper, a generic Bayesian framework is proposed for training context-dependent classification rules from wide-area airborne LWIR imagery. The proposed framework combines sparse classification models with either supervised or discriminative context identification to pool information across contexts and improve classification overall. Experiments are performed with data from a LWIR landmine detection system. Contexts are learned from endmember abundances extracted from the background near each detected anomaly. Classification performance is compared with single-classifier approaches using the same information as well as other baseline anomaly detectors from the literature. Results indicate that utilizing context for classifying anomalies in HSI could lead to more robust performance over varying terrain.

Ground-penetrating radar (GPR) is a powerful and rapidly maturing technology for subsurface threat identification. However, sophisticated processing of GPR data is necessary to reduce false alarms due to naturally occurring subsurface clutter and soil distortions. Most currently fielded GPR-based landmine detection algorithms utilize feature extraction and statistical learning to develop robust classifiers capable of discriminating buried threats from inert subsurface structures. Analysis of these techniques indicates strong underlying similarities between efficient landmine detection algorithms and modern techniques for feature extraction in the computer vision literature. This paper explores the relationship between and application of one modern computer vision feature extraction technique, namely histogram of oriented gradients (HOG), to landmine detection in GPR data. The results presented indicate that HOG features provide a robust tool for target identification for both classification and prescreening and suggest that other techniques from computer vision might also be successfully applied to target detection in GPR data.

Frequency-domain electromagnetic induction (EMI) sensors can measure object-specific signatures that can be used to discriminate landmines from harmless clutter. In a model-based signal processing paradigm, the object signatures can often be decomposed into a weighted sum of parameterized basis functions, such as the discrete spectrum of relaxation frequencies (DSRF), where the basis functions are intrinsic to the object under consideration and the associated weights are a function of the target-sensor orientation. The basis function parameters can then be used as features for classifying the target. One of the challenges associated with effectively utilizing a model-based signal processing paradigm such as this is determining the correct model order for the measured data, as the number of basis functions containing fundamental information regarding the target under consideration is not known a priori. In this paper, sparse Bayesian relevance vector machine (RVM) regression is applied to simultaneously determine both the number of parameterized basis functions and their relative contributions to the measured signal assuming a DSRF signal model. The target is then classified utilizing the basis function parameters as features within a statistical classifier. Results for data measured with a prototype frequency-domain EMI sensor at a standardized test site are presented, and indicate that RVM regression followed by distance-based statistical classifiers utilizing the resulting model-based features provides an effective approach for classifying and identifying landmine targets.

Detection of buried explosive threats is a challenging problem. GPR has recently become a powerful tool for achieving robust subsurface target detection, but novel target types, and large numbers of subsurface objects in rural environments significantly complicate accurate discrimination of explosive threats from harmless false alarms. Significant research in feature extraction from GPR data has previously shown the capability for improved performance. Similarly, many techniques from the computer vision literature have made significant strides in recent years in for improvements in object class recognition. This work studies the relationships between and application of feature descriptor techniques from the computer vision community in application to target detection in GPR data. Relationships between a very successful computer vision technique (Histogram of Oriented Gradients) and a related powerful technique from subsurface sensing (Edge Histogram Descriptors) are explored, and preliminary results suggest that techniques from the computer vision literature may provide robust target detection performance in GPR.

In recent years, context-dependent algorithm fusion has been proposed for improving landmine detection with ground penetrating radar (GPR) across changing environmental and operating conditions. While context-dependent fusion techniques generally assume independent observations, previous work showed that spatial information may be exploited by modeling context with a hidden Markov model (HMM). However, the degree of performance improvement was found to depend the number of states included in the HMM. In this work, stick-breaking priors were employed to automate learning of the number of HMM states, and therefore the number of contexts to consider. The improved spatially-dependent fusion technique was evaluated on GPR data collected over various targets at multiple test sites, and performance was compared to another context-dependent technique which assumed independent observations. Results illustrate the potential for nonparametric, spatially-dependent context modeling to exploit contextual information in sequentially-collected GPR data and improve overall classification performance.

A technique for making landmine detection with a ground-penetrating radar (GPR) sensor more robust to fluctuations in environmental conditions is presented. Context-dependent feature selection (CDFS) counteracts environmental uncertainties that degrade detection and discrimination performances by modifying decision rules based on inference of the environmental context. This paper utilized both physics-based and statistical methods for extracting features from GPR data to characterize surface texture and subsurface electrical properties, and a nonparametric hypothesis test was used to identify the environmental context from which the data were collected. The results of probabilistic context identification were then used to fuse an ensemble of classifiers for discriminating landmines from clutter under diverse environmental conditions. CDFS was evaluated on a large set of GPR data collected over several years in different weather and terrain conditions. Results indicate that our context-dependent technique improved landmine discrimination performance over conventional fusion of several currently fielded algorithms from the recent literature.

A framework is presented for information-theoretic sensor management for the detection of static targets. The sensor manager searches for targets within a cell grid using a suite of sensor platforms. Each sensor platform may contain one or more sensing modalities, and each of these modalities has known probabilities of detection and false alarm and also has an associated cost of use. Additional information such as motion constraints on the sensors and the prior distribution of the targets in space is incorporated. The sensor manager then directs the movement of the sensors through the grid by maximizing the expected information gain that will be obtained with each new sensor observation. Key modeling questions are addressed, including the selection of an appropriate information measure and the joint or independent management of the sensors. Through a number of simulations, the performance of the sensor manager is compared to the performance of a blind sweep procedure, a random search procedure, and an alternative information-theoretic sensor manager. The intelligent sensor management procedure is demonstrated to achieve a superior performance compared to all of the other three techniques. A specific application area for which the sensor management problem is becoming more critical is landmine detection; thus, the performance of the sensor manager is also analyzed using real data from three different landmine detection sensing modalities, and the proposed sensor management technique is again demonstrated to be superior compared to more simplistic approaches.

It is not uncommon for cochlear implant patients to have individual electrodes that produce anomalous percepts that impair or prevent the effective transmission of auditory information. Exhaustive psychophysical testing to detect all such information channels is time prohibitive; however, if impaired channels could be identified quickly then the application of remediation strategies becomes more cost-effective. Previous studies have suggested that the missing speech information could produce a predictable pattern of errors in speech-token identification tasks. This study investigates the application of partial least squares discriminant analysis to identifying the presence of channel-specific impairments based on confusion matrices generated from vowel and consonant token identification tasks. The results of this study, using normal-hearing subjects tested with acoustic models, suggest that the partial least squares discriminant analysis can successfully distinguish impaired and unimpaired models, as well as identify channel-specific impairments, without requiring a significant amount of labeled training data.

Context-dependent approaches to landmine detection have been developed in recent years to exploit the sensitivity of ground-penetrating radar (GPR) to changes in environmental conditions. Previous approaches to context-dependent fusion have only considered the special case of statistically independent observations. This work proposes the use of Hidden Markov Models, trained on the GPR background, for modeling the context of observation sequences. The performances of context-dependent fusion using two statistical context models were compared in an experiment with field data. One approach utilized a Hidden Markov Context Model (HMCM), and the other utilized a Gaussian mixture. Experimental results illustrated that the HMCM improved performance of context-dependent fusion. These results suggest that spatial dependencies are an important source of contextual information for landmine detection that warrants further investigation.

Many of the ongoing efforts to develop strategies for detecting and locating subsurface unexploded ordnance (UXO) use features based on phenomenological models to discriminate between UXO and harmless clutter. The process of generating features requires model inversion to fit the phenomenological model to the measured sensor data. In commonly-used model inversion processes, the standard measures of model fit error do not incorporate the spatial distribution of the data used in the model inversion. This study incorporates the Fisher information in a joint metric optimization to assess the spatial distribution of data and how well the model parameters are supported by the data used in the model inversion. The outcomes of this study indicate that some outliers in the feature space can be mitigated by considering the Fisher information in the model inversion process, resulting in improved unexploded ordnance detection rates in a test using data collected at Camp Sibert, Alabama.

Ground penetrating radar (GPR) is a promising technique for buried threat detection which provides a complimentary phenomenology to electro-magnetic induction (EMI) based sensing. However, many successful GPR-based buried threat detection algorithms require data collected both before and after an object of interest is encountered to make a declaration (typically this data is used to perform background normalization, or to adequately characterize the object's shape). Samples taken past an object of interest, but before a decision is made, constitute an algorithm's “spatial latency”. For vehicular mounted antennae arrays, where vehicle stopping distance is a function of vehicle dynamics, driver responsiveness, and algorithmic spatial latency, reducing an algorithm's spatial latency can increase overall system safety and help keep operators out of harm's way. In this work we propose a stochastic sampling algorithm that can help reduce spatial latency for a wide range of GPR-based buried threat detection algorithms.

Previously, a framework for sensor management has been developed for the detection of static targets such as landmines. The sensor manager functions by tasking the available sensors to greedily maximize the expected information gain obtained with each new sensor observation. This paper examines several of the key assumptions and decisions that were made in the formulation of this sensor manager to assess both the validity and the performance effects of these decisions. Specifically, this paper examines which unmanaged sensing technique is best used to make performance comparisons with the sensor manager, whether multiple sensors should be optimized independently or jointly, and finally whether the Kullback-Leibler divergence or Renyi divergence is the best choice of information measure to use. This paper demonstrates that in all three cases, the choices made in the original formulation of the sensor manager are the most effective and appropriate.

A method of generating features for classification of unexploded ordnance using matching pursuits decomposition is proposed as a possible alternative, or complement, to the use of a standard dipole model. The proposed matching pursuit decomposition with an application-specific dictionary was evaluated using measured signatures from both isolated anomalies and overlapping signatures interpolated from multiple, isolated anomalies. Results indicate that matching pursuits decomposition provides computationally inexpensive and consistent representations of the anomalies in feature space, suggesting the potential utility of such an approach for UXO classification.

As ground penetrating radar phenomenology continues to improve, more advanced statistical signal processing approaches become applicable to subsurface inference in GPR data. Despite the wide body of literature exploring the applications of various approaches to processing GPR data, statistical modeling of realistic soil responses is a difficult task, and the algorithms developed for real-time fielded GPR processing are rarely directly motivated by statistical models of GPR data. In this work, we present a tractable spatial statistical model for volumetric GPR data which can be used to motivate the application of various signal processing approaches to solving problems of interest in GPR data like pre-screening, feature extraction, and air/ground response tracking.

Phenomenological modeling coupled with statistical signal processing has been shown to significantly improve capabilities for discriminating unexploded ordnance (UXO) from benign clutter using electromagnetic induction (EMI) sensor data. The general premise underlying the majority of these coupled approaches is that a phenomenological model is fit to the measured data, and the parameters estimated from this model inversion, which characterize the interrogated target, are utilized in subsequent statistical signal processing algorithms to classify the target as either UXO or clutter. A potential limitation of this coupled approach is that the inversion has been shown to be sensitive to uncertainty associated with the sensor positions. When the measurement positions are uncertain, the inversion results are more variable, and consequently, discrimination performance degrades. In this letter, a Bayesian methodology is applied to estimate the desired features from the measured data. This method explicitly acknowledges that uncertainty in the sensor positions exists and incorporates this knowledge to find the maximum-likelihood feature estimates by integrating over the uncertain measurement positions. Due to the high dimensionality of the integration, Monte Carlo integration, a statistical technique to estimate the value of an integral, is employed. Simulation results show that this Bayesian approach in mitigating sensor position uncertainty produces features with lower variability and, therefore, provides improved discrimination performance.

Cochlear implant patients exhibit a wide range of performance on speech recognition tasks. One potential explanation for such variability is the existence of psychophysically observed phenomena that might indicate the presence of anomalous percepts associated with certain electrical stimuli, which in turn could limit the transmission of important auditory cues. Exhaustive psychophysical testing to detect all such psychophysical anomalies is time prohibitive; however, the search for anomalous channels could be expedited with prior information identifying channels potentially containing an anomaly. This study proposes a method of analyzing confusion matrices from speech token recognition tasks with the intent of identifying impaired channels. Results using both normal-hearing subjects tested with impaired acoustic models and cochlear implant subjects suggest that the proposed methods are providing information about the probability of impairment on each channel.

Numerous detection algorithms, using various sensor modalities, have been developed for the detection of mines in cluttered and noisy backgrounds. The performance for each detection algorithm is typically reported in terms of the receiver operating characteristic (ROC), which is a plot of the probability of detection versus false alarm as a function of the threshold setting on the output decision variable of each algorithm. In this paper, we present multisensor decision-fusion algorithms that combine the local decisions of existing detection algorithms for different sensors. This offers an expedient, attractive, and much simpler alternative to the design of an algorithm that fuses multiple sensors at the data level, especially in cases of limited training data where it is difficult to make accurate estimates of multidimensional probability density functions. The goal of our multisensor decision-fusion approach is to exploit the complimentary strengths of existing multisensor algorithms so as to achieve performance (ROC) that exceeds the performance of any sensor algorithm operating in isolation. Our approach to multisensor decision fusion is based on the optimal signal detection theory using the likelihood ratio. We consider the optimal fusion of local decisions for two sensors: a ground-penetrating radar and a metal detector. A new robust algorithm for decision fusion that addresses the problem in which the statistics of the training data are not likely to exactly match the statistics of the test data is presented. ROCs are presented and compared for field data

In this paper, we consider the application of texture features for antitank landmine detection in ground- penetrating-radar data in the difficult scenario of very high clutter environments. In particular, we develop a technique for 3-D texture feature extraction, and we compare the results for landmine/clutter discrimination using classifiers that are built on 3-D as well as on 2-D texture feature sets. Our results indicate performance improvements across several different challenging testing scenarios when using the relevance-vector-machine classifiers that are trained on our 3-D feature sets as compared to the performance using the 2-D texture feature sets.

Recent advances in ground penetrating radar (GPR) fabrication and signal processing have made high fidelity detection of buried anti-tank landmines a practical possibility under field scenarios. However, detection of subsurface landmines at a low false alarm rate (FAR) requires the effective removal of the response from the air/ground interface (ground-bounce response). This in turn requires accurate and automatic tracking of the time of arrival of the air/ground interface in time-domain GPR data. Such tracking of the ground bounce response can be difficult to perform under certain conditions including the presence of surface-laid landmines, surface vegetation, snow drifts, and multiple subsurface structures like buried roadbeds. In this work, we will explore the application of a low- latency Kalman filter applied to ground bounce tracking in GPR data and resulting performance improvements for pre-screening algorithms under extreme weather conditions.

Ground-penetrating radar (GPR) has been widely used for landmine detection due to its high signal-to-noise ratio (SNR) and superior ability to image nonmetallic landmines. Processing GPR data to obtain better target images and to assist further object detection has been an active research area. Phase-shift migration is a widely used method; however, its wavenumber space is nonuniformly sampled because of the nonlinear relationship between the uniform frequency samples and the wavenumbers. Conventional methods use linear interpolation to obtain uniform wavenumber samples and compute the fast Fourier transform (FFT). This paper develops two- and three-dimensional migration methods that process GPR data to obtain images close to the actual target geometries using a nonuniform fast Fourier transform (NUFFT) algorithm. The proposed method is first compared to the conventional migration approaches on simulated data and then applied to landmine field data sets. Results suggest that the NUFFT migration method is useful in focusing images, estimating landmine structure, and retaining relatively high signal-to-noise ratio in the migrated data. The processed data sets are then fed to the normalized energy and least-mean-square-based anomaly detectors. Receiver operating characteristic curves of data sets processed by different migration methods are compared. The NUFFT migration shows potential improvements on both classifiers with a reduced false alarm rate at most probabilities of detection.

Quadrupole resonance (QR) is a novel technology recently applied to landmine detection. The detection process is specific to the chemistry of the explosive, and therefore is less susceptible to the types of false alarms experienced by metal detectors and ground-penetrating radars. Although QR is vulnerable to radio-frequency interference (RFI) when the sensor is deployed in the field, adaptive RFI mitigation can remove most of the RFI. In this paper, advanced signal processing algorithms applied to the postmitigation signal are studied to enhance explosive detection. A new Kalman filtering strategy is proposed to estimate and detect the QR signal in the postmitigation signal. The results using both simulated data and experimental data show that the proposed algorithm can provide robust landmine detection performance.

Ground-penetrating radar (GPR) has been proposed as an alternative to classical electromagnetic induction techniques for the landmine detection problem. The Wichmann/Niitek system provides a good platform for novel GPR-based antitank mine detection and classification algorithm development due to its extremely high SNR. When the GPR sensor is mounted on a moving vehicle, the target signatures are hyperbolas in a time-domain data record. The goal of this work is to extract useful features that exploit this knowledge in order to improve target detection. The algorithms can be divided into two steps: feature extraction and classification. Preprocessing is also considered to remove both stationary effects and nonstationary drift of the data and to improve the contrast of the desired hyperbolas. The algorithm is evaluated using real data over primarily plastic antitank mines collected with a fielded GPR sensor at a government test site.

Cochlear implants are able to restore some degree of hearing to deafened individuals; however implant users are particularly susceptible to background noise. The effect of noise can be assessed using vowel and consonant confusions measured in listening experiments. The paper presents three signal processing methods developed to predict patterns in vowel and consonant confusion in noise for cochlear implant users. Prediction performance is tested using the results of a listening experiment conducted with acoustic models of two cochlear implant speech processors and normal hearing subjects. Confusion prediction is based on prediction metrics calculated using each method's unique representation of the speech tokens.

We investigate dynamic range and intensity discrimination for electrical noise-modulated pulse-train stimuli using a stochastic auditory nerve model (Bruce, I.C. et al., IEEE Trans. Biomed Eng., vol.46, no.6, p.617-29, 1999). Based on a hypothesized monotonic relationship between loudness and the number of spikes, the theoretical prediction of the most uncomfortable level was determined by comparing spike counts to a fixed threshold (Bruce et al., IEEE Trans. Biomed Eng., vol.46, no.6, p.1393-1404, 1999). However, no specific rule for determining this fixed number has previously been suggested. We determine the most uncomfortable level based on the excitation pattern of the basilar membrane in a normal ear. The number of fibers corresponding to the portion of the basilar membrane driven at an uncomfortable stimulus level in a normal ear is related to the most uncomfortable spiking number. The intensity discrimination limens are predicted using signal detection theory via the probability mass function (PMF) of the neural response and via experimental simulations. The results show that the uncomfortable level for a pulse-train stimulus increases slightly as noise level increases. Combining this with our previous threshold predictions (Xu and Collins, IEEE Trans. Biomed. Eng.), we hypothesize that the dynamic range for noise-modulated pulse-train stimuli increases with additive noise. However, since our predictions indicate that intensity discrimination under noise degrades, the overall intensity coding performance does not improve significantly.

Recent advances in ground penetrating radar (GPR) fabrication and algorithm development have yielded significant performance improvements for anti-tank landmine detection in government sponsored blind tests. However, these blind tests are typically conducted over well-maintained homogeneous testing lanes specifically designed to test landmine detection performance in low-clutter population situations. New GPR data collections over targets emplaced in un-maintained off-lane soils have much higher GPR anomaly populations and provide more stringent tests of landmine detection algorithms. In this work, we focus on the application of feature-based class separation techniques to lower false alarm rates in heterogeneous off-road soils. In particular, we explore the application of texture feature coding methods (TFCM), which have previously shown promise in fields like tumor detection

Quadrupole resonance (QR) technology for the detection of explosives is of crucial importance in an increasing number of applications. For landmine detection, where the detection system cannot be shielded, QR has proven to be highly effective if the QR sensor is not exposed to radio-frequency interference (RFI). However, strong non-Gaussian RFI in the field is unavoidable. A statistical model of such non-Gaussian RFI noise is given in this letter. In addition, the asymptotic Cramer-Rao lower bound for estimating a deterministic QR signal in this non-Gaussian noise is presented. The performance of several convenient estimators is compared to this bound.

Interpretation of psychophysical data from impaired-hearing individuals on intensity discrimination tasks has been confounded by the fact that some impaired individuals' performance is near-normal in quiet, whereas for others, the difference limen is elevated. It has been observed that a subject's discrimination abilities may be related to the underlying audiogram configuration, which is often dependent on the type of physiological damage that has occurred. This suggests that data be grouped and analyzed according to the type of hearing loss. An experimental study by Schroder et al. (1994) tested the hypothesis that the near-normal performance of some impaired subjects was the result of a normal spread of excitation, rather than greater intensity resolution due to loudness recruitment. In this paper, we replicate the trends observed in Schroder et al.'s experimental data using a combination of signal detection theory and a computational auditory model. By linking simulated psychophysical predictions with the corresponding simulated physiological responses, this theoretical analysis provides further qualitative support for the hypothesis that the observed performance is due to the spread of excitation.

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.

Previous work in subsurface object discrimination using electromagnetic induction data has shown that discrimination algorithms based on statistical signal processing techniques are effective for classifying data from objects that occur in isolation. However, for multiple closely spaced subsurface objects, the raw (unprocessed) measurement is a mixture of the responses from several objects and as such cannot be used directly to determine the identity of each of the individual objects. Thus, we propose to separate individual signatures from the mixture by posing the problem as a blind source separation (BSS) problem and effecting signature separation using independent component analysis. We propose to apply BSS to separate the mixed signatures and then follow the separation process with a Bayesian classifier. This approach is evaluated using both simulated data and data from unexploded ordnance items. The results show that this approach can be used to effectively classify multiple closely spaced objects.

Decay rate estimation has been proposed as an effective method for signal characterization in many application areas, including subsurface sensing using time-domain electromagnetic induction (EMI) sensors. The physical basis for this strategy is that every signal of interest, which corresponds to a target or phenomenon of interest, possesses a unique set of decay rates. In theory, the characteristic decay rates can be estimated from the measured signal, and then utilized for signal detection and subsequent identification. Using this approach, signal discrimination performance is dependent upon decay rate estimation performance. The Cramer-Rao lower bound (CRLB) for decay rate and amplitude coefficient estimates is utilized to investigate the fundamental limits of decay rate estimation accuracy. Previous derivations of the CRLB for decay rate estimates have focused on signals which are linearly sampled beginning at time t = 0. Here, the CRLB is generalized to accommodate any arbitrary sampling method and any initial starting time. A parameter transformation which improves decay rate estimation performance is also presented. Simulation results across a wide range of decay rates and SNRs show that nonlinear least squares estimation of the decay rates via the proposed transformation provides estimates with smaller RMS and bias than can be obtained without the parameter transformation. The parameter transformation also provides decay rate estimates that approach the CRLB. Improvement in estimation performance for this class of signals has important ramifications in signal detection, classification, and identification performance in several geophysical application areas.

We consider the detection of subsurface unexploded ordnance via magnetometer and electromagnetic-induction (EMI) sensors. Detection performance is presented, using model-based signal processing algorithms. We first develop and validate the parametric models, using both numerical and measured data. These models are then applied in the context of feature extraction, and the features are processed via two signal-processing algorithms. The detection algorithms are discussed in detail, with comparisons made based on performance with measured magnetometer and EMI data.

The effect of cochlear hearing loss on intensity discrimination has been experimentally investigated and reported in the literature. It has been observed that the impairment configuration has a significant effect on a subject's discrimination abilities. In some cases, intensity discrimination for an impaired individual is near-normal at equal SPLs, whereas in other cases, the difference limen is elevated. It has been hypothesized that the near-normal performance of some impaired individuals may be due to either greater intensity resolution (resulting from loudness recruitment) or to the normal spread of excitation. In this paper, we simulate an experiment conducted by Schroder et al. (1994) devised to test these hypotheses. Using a signal detection theory-based approach in combination with a computational auditory model, we are able to replicate their experimental results. Additionally, by manipulating the model, we are able to demonstrate theoretically that the observed behavior results from the spread of excitation.

The effect of a low level additive noise process on the input and output characteristics and threshold behavior of auditory nerves (ANs) is studied by means of a stochastic computational model. This paper derives the stochastic properties of the model input and output for adaptive threshold procedures. A closed form solution for the input, or amplitude, probability distribution is obtained via Markov models. The output statistics are derived by integrating over the noise-free probability mass function (PMF). All theoretical PMFs are verified by simulations. Theoretical threshold predictions as a function of noise level are made based on these PMFs and the results indicate that threshold is adversely affected by the presence of low levels of noise.

A new adaptive time delay estimation method is proposed for highly corruptive environments based on the cyclostationary property of the source signal. The time delay operator is modeled as a finite impulse response filter. The new adaptive scheme is based on parametric modeling between two sensor measurements and employs cyclic statistics of the data. Simulation examples are presented to demonstrate the signal selectivity of the new adaptive method in the presence of noises and interference. The only requirement is that the source signal has a known (or measurable) carrier frequency or keying rate that is distinct from those of interfering signals.

Detection and remediation of unexploded ordnance (UXO) represents a major challenge on closed, closing, and transferred military ranges as well as on active installations. The detection problem is exacerbated by the fact that on sites contaminated with UXO, extensive surface and sub-surface clutter and shrapnel is also present. Traditional methods used for UXO remediation have difficulty distinguishing buried UXO from these anthropic clutter items as well as from naturally occurring magnetic geologic noise, and thus incur prohibitively high false alarm rates. The reduction of the false alarm rate has proven to be the greatest challenge for UXO remediation. In this paper, sensor fusion techniques are applied to field data from magnetometer and electromagnetic induction (EMI) sensors in order to determine to what degree such an approach results in false alarm mitigation. The adoption of a model consisting of multiple non-colocated dipoles is shown to improve our ability to predict measured signatures. The results indicate that performance can be improved by limiting the processing bandwidth to those frequencies that are the most robust to naturally occurring geological noise.

Nuclear quadrupole resonance (NQR) technology for the detection of explosives is of crucial importance in an increasing number of applications. For landmine detection, NQR has proven to be highly effective if the NQR sensor is not exposed to radio frequency interference (RFI). Since strong nonstationary RFI in the field is unavoidable, a robust detection method is required. With the aid of reference antennas, a frequency domain LMS algorithm is applied to cancel the RFI in field data. An average power detector based on power spectral estimation algorithms is proposed and performance using both the periodogram and MUSIC algorithms is evaluated. The detection performance has been compared with that of a non-adaptive Bayesian detector. The experimental results show that, unlike the non-adaptive Bayesian detector, the average power detector provides perfect detection capability if the data segments involved in the collection process are sufficiently long.

Weighted sums of decaying exponentials characterize the response of many physical systems. Therefore, accurate decay rate estimation is a goal in many diverse disciplines. A parameter transformation which improves decay rate estimation is presented. Simulation results across a wide range of decay rates, signal-to-noise ratios, and ratios of decay rates show that nonlinear least squares estimation of the decay rates via the proposed parameter transformation provides estimates with smaller RMS errors and bias than can be obtained without the parameter transformation. In addition, it is shown that the parameter transformation provides decay rate estimates which are closer to achieving the Crame/spl acute/r-Rao bound. Improvement in estimation performance for this class of signals has important ramifications in signal detection performance in several application areas.

The response of time-domain electromagnetic induction (EMI) sensors, which have been used almost exclusively for landmine detection, is related to the amount of metal present in the object and its distance from the sensor. Unluckily, there is often a significant amount of metallic clutter in the environment that also induces an EMI response. Consequently, EMI sensors employing detection, algorithms based solely on metal content suffer from large false alarm rates. To mitigate this false alarm problem for mines with substantial metal content, statistical algorithms have been developed that exploit models of the underlying physics. In such models it is commonly assumed that the soil has a negligible effect on the sensor response, thus the object is modeled in "free space." We report on studies that were performed to test, the hypotheses that for broadband EMI sensors: 1) soil cannot be modeled as free space when the buried object has low metal content and 2) advanced signal processing algorithms can be applied to reduce the false alarm rates. Our results show that soil cannot be modeled as free space and that when modeling soil correctly our advanced algorithms reduced the false alarm probability by up to a factor of 10 in blind tests.

We have investigated the differences between normal and impaired auditory processing for a frequency discrimination task by analyzing the responses of a computational auditory model using signal detection theory. Two detectors, one using all of the information in the signal, the other using only the number of neural responses, were implemented. An evaluation of the performance differences between the two theoretical detectors and experimental data may provide insight into quantifying the type of information present in the auditory system as well as whether the human auditory system uses this information efficiently. Results support previous hypotheses that, for lowand mid-range frequencies, the auditory system is able to use temporal information to perform frequency discrimination (see Moore, B.C.J., J. Acoust. Soc. Am., vol.54, p.610-19, 1973). The results also suggest that some temporal information is represented in the neural spike train, even at high frequencies However, the ability of the auditory system to use this information deteriorates at higher frequencies.

We focus on the development of signal processing algorithms that incorporate the underlying physics characteristic of the sensor and of the anticipated unexploded ordnance (UXO) target, in order to address the false alarm issue. In this paper, we describe several algorithms for discriminating targets from clutter that have been applied to data obtained with the multisensor towed array detection system (MTADS). This sensor suite includes both electromagnetic induction (EMI) and magnetometer sensors. We describe four signal processing techniques: a generalized likelihood ratio technique, a maximum likelihood estimation-based clustering algorithm, a probabilistic neural network, and a subtractive fuzzy clustering technique. These algorithms have been applied to the data measured by MTADS in a magnetically clean test pit and at a field demonstration. The results indicate that the application of advanced signal processing algorithms could provide up to a factor of two reduction in false alarm probability for the UXO detection problem.

The performance of subsurface target identification algorithms using data from time-domain electromagnetic induction (EMI) sensors is investigated. The response of time-domain EMI sensors to the presence of a conducting object may be modeled as a weighted sum of decaying exponential signals. Although the weights associated with each of the modes are dependent on the target/sensor orientation, the decay rates are a function of the target's composition and geometry and therefore are intrinsic to the target. Since the decay rates are not dependent on target/sensor orientation or other unobservable parameters, decay rate estimation has previously been proposed as a viable method for target identification. The performance attained with Bayesian target identification algorithms operating on the entire time-domain signal and decay rate estimates is compared through both numerical simulations and application to experimental data. The decay rate estimates utilized in the numerical simulations are assumed to achieve the Cramer-Rao lower bound (CRLB), which provides a lower bound on the variance of an unbiased parameter estimate. The simulations as well as results obtained with experimental data show that processing the entire time-domain signal provides better target identification and discrimination performance than processing decay rate estimates.

The physical phenomenology of electromagnetic induction (EMI) sensors' is reviewed for application to land mine detection and remediation. The response from time-domain EMI sensors is modeled as an exponential damping as a function of time, characterized by the initial magnitude and decay rate. Currently deployed EMI sensors that are used for the land mine detection process the recorded signal in a variety of ways in order to provide an audio output for the operator to judge whether or not the signal is from a mine. Sensors may sample the decay curve, sum it, or calculate its energy. Based on exponential decay model and the assumption that the sensor response is subject to additive white Gaussian noise, the performance of these, as well as optimal, detectors are derived and compared. Theoretical performance predictions derived using simplifying assumptions are shown to agree closely with simulated performance. It will also be shown that the generalized likelihood ratio test (GLRT) is equivalent to the likelihood ratio test (LRT) for multichannel time-domain EMI sensor data under the additive white Gaussian noise assumption and specific assumptions regarding the statistics of the decay rates of targets and clutter.

In their previous work, the authors have shown that the detectability of landmines can be improved dramatically by the careful application of signal detection theory to time-domain electromagnetic induction (EMI) data using a purely statistical approach. In this paper, classification of various metallic land-mine-like targets via signal detection theory is investigated using a prototype wideband frequency-domain EMI sensor. An algorithm that incorporates both a theoretical model of the response of such a sensor and the uncertainties regarding the target/sensor orientation is developed. This allows the algorithms to be trained without an extensive data collection. The performance of this approach is evaluated using both simulated and experimental data. The results show that this approach affords substantial classification performance gains over a standard approach, which utilizes the signature obtained when the sensor is centered over the target and located at the mean expected target/sensor distance, and thus ignores the uncertainties inherent in the problem. On the average, a 60% improvement is obtained.

Our previous work has indicated that the careful application of signal detection theory can dramatically improve detectability of landmines using time-domain electromagnetic induction (EMI) data. In this paper, classification of various metal targets via signal detection theory is investigated using a prototype wideband frequency-domain EMI sensor. An algorithm that incorporates both the uncertainties regarding the target-sensor orientation and a theoretical model of the response of such a sensor is developed. The performance of this approach is evaluated using both simulated and experimental data. The results show that this approach affords substantial classification performance gains over the traditional matched filter approach, on average by 60%.

In order to develop improved remediation techniques for hearing impairment, auditory researchers must gain a greater understanding of the relation between the psychophysics of hearing and the underlying physiology. One approach to studying the auditory system has been to design computational auditory models that predict neurophysiological data such as neural firing rates. To link these physiologically-based models to psychophysics, theoretical bounds on detection performance have been derived using signal detection theory to analyze the simulated data for various psychophysical tasks. Previous efforts, including our own recent work using the auditory image model, have demonstrated the validity of this type of analysis; however, theoretical predictions often exceed experimentally-measured performance. In this paper, we compare predictions of detection performance across several computational auditory models. We reconcile some of the previously observed discrepancies by incorporating phase uncertainty into the optimal detector.

Classification of metal objects is important for landmine and unexploded ordnance applications. Previously, we have in investigated optimal classification of landmine-like metal objects using wideband frequency-domain electromagnetic induction data. Here, a suboptimal processor, which is computationally less burdensome than the optimal processor, is discussed. The data is first normalized, exploiting the fact that the level of the response changes significantly while the structure of the magnitude of the response changes only slightly as the target/sensor orientation changes for the class of objects considered. Results indicate that the suboptimal processor performance approaches that of the optimal classifier on normalized data. Thus, normalization mitigates the uncertainty resulting from the target/sensor orientation.

A rigorous signal detection theoretic analysis is used to improve detectability of land mines. The development is performed for sensors that integrate time-domain information to provide a single data point (standard metal detector), those that provide a sampled portion of the time-domain waveform, and those that operate at several discrete frequencies. This approach is compared to standard thresholding techniques, and it is shown to provide substantial improvements when evaluated on measured data.

We demonstrate the utility of a multiresolution approach for target detection in SAR imagery. Man-made objects exhibit characteristic phase and amplitude fluctuations as the image resolution is varied, while natural terrain has a random signature. We construct a number of detection strategies: an optimal invariant multiresolution detector based on a derived multiresolution increments process; and a generalized likelihood ratio detector to differentiate between a first-order autoregressive multiresolution increments process and white noise. We show that these schemes significantly outperform a standard energy detector operating on the finest available SAR resolution.

Examines the problem of detecting a known target in clutter and shows that a multiresolution-based detector significantly outperforms a more conventional single-resolution detector. The authors then apply a sampling strategy that allows one to choose, for each fixed n, those n resolutions that maximize detection performance. This optimal multiresolution sampling strategy suggests that wavelet or wavelet packet decompositions which typically use dyadic or triadic resolution splits are not necessarily optimal bases for coherent radar signature representation.>