Over the last two decades, researchers investigated various approaches for detection and classification of targets in forward looking infrared (FLIR) imagery using correlation based techniques. In this paper, a novel technique is proposed to recognize and track single as well as multiple identical and/or dissimilar targets in real life FLIR sequences using a combination of extended maximum average correlation height (EMACH) and polynomial distance classifier correlation filter (PDCCF). The EMACH filters are used for the detection stage and PDCCF filter is used for the classification stage for improving the detection and discrimination capability. The EMACH and PDCCF filters are trained a priori using target images with expected size and orientation variations. In the first step, the input scene is correlated with all the detection filters (one for each desired or expected target class) and the resulting correlation outputs are combined. The regions of interest (ROI) are selected from the input scene based on the regions with higher correlation peak values in the combined correlation output. In the second step, PDCCF filter is applied to these ROIs to identify target types and reject clutters/backgrounds based on a distance measure and a threshold. Moving target detection and tracking is accomplished by applying this technique independently to all incoming image frames. Independent tracking of target(s) from one frame to the other allows the system to handle complicated situations such as a target disappearing in few frames and then reappearing in later frames. This method has been found to yield robust performance for challenging FLIR imagery in terms of faster and accurate detection and classification as well as tracking of the targets.

Automatic target recognition, or target discrimination, is an ever-increasing need in both tactical and strategic engagements. Complex imaging systems, such as those based on adaptive optics compensation, are proven but expensive. This paper addresses a specific problem amenable to non-imaging distinction of targets, specifically symmetric and oblong. The approach is to illuminate a target with a rotating oblong beam, either in the near field (tactical) or far field (strategic). The returns from a symmetric object in the absence of pointing errors will be constant while returns from an oblong object will produce a sinusoidal signal, thus distinguishing the objects without imaging. This paper addresses the result of illumination with an oblong beam, starting with a pure beam, a beam corrupted by atmospheric effects and a beam affected by pointing errors referred to as jitter and boresight.

A shape-based algorithm, designated SPOT, is applied to the output of ARL’s, contrast-like feature based, FLIR anomaly target detector in an effort to improve clutter rejection. The shape-based algorithm uses a spatial symmetry operator that discriminates man-made structures in imagery. The method uses no predetermined templates or filters, which would be range and/or aspect dependent, but rather generates templates on-the-fly from the input data. Results explore the application of one such symmetry operator, and present a comparative analysis of target detection performance, based on ROC curves and detection histograms.

Despite the effort made in the eye tracking community, the eye position tracking problem remains unsolved completely, due to the large variations in the eye appearance. This paper describes a multi-modal eye position tracker in dark/bright pupil image sequences. The tracking algorithm consists of detecting meaningful particles that correspond to IR-Pupil responses and eye motion, altering of particles through appearance models in the local invariant descriptor space, and matching of eye neighbors. Experimental validations have shown satisfactory performance in term of precision of eye position estimation, and robustness to 2D head rotations, translations, closed eye states, and reasonable out-of-plane rotations.

Following the tendency of increased use of imaging sensors in military aircraft, future fighter pilots will need onboard artificial intelligence e.g. ATR for aiding them in image interpretation and target designation. The European Aeronautic Defence and Space Company (EADS) in Germany has developed an advanced method for automatic target recognition (ATR) which is based on adaptive neural networks. This ATR method can assist the crew of military aircraft like the Eurofighter in sensor image monitoring and thereby reduce the workload in the cockpit and increase the mission efficiency. The EADS ATR approach can be adapted for imagery of visual, infrared and SAR sensors because of the training-based classifiers of the ATR method. For the optimal adaptation of these classifiers they have to be trained with appropriate and sufficient image data. The training images must show the target objects from different aspect angles, ranges, environmental conditions, etc. Incomplete training sets lead to a degradation of classifier performance. Additionally, ground truth information i.e. scenario conditions like class type and position of targets is necessary for the optimal adaptation of the ATR method. In Summer 2003, EADS started a cooperation with Kongsberg Defence & Aerospace (KDA) from Norway. The EADS/KDA approach is to provide additional image data sets for training-based ATR through IR image simulation. The joint study aims to investigate the benefits of enhancing incomplete training sets for classifier adaptation by simulated synthetic imagery. EADS/KDA identified the requirements of a commercial-off-the-shelf IR simulation tool capable of delivering appropriate synthetic imagery for ATR development. A market study of available IR simulation tools and suppliers was performed. After that the most promising tool was benchmarked according to several criteria e.g. thermal emission model, sensor model, targets model, non-radiometric image features etc., resulting in a recommendation. The synthetic image data that are used for the investigation are generated using the recommended tool. Within the scope of this study, ATR performance on IR imagery using classifiers trained on real, synthetic and mixed image sets was evaluated. The performance of the adapted classifiers is assessed using recorded IR imagery with known ground-truth and recommendations are given for the use of COTS IR image simulation tools for ATR development.

In this paper, we investigate the detection and classification of targets in forward-looking infrared (FLIR) imagery under various challenging scenarios. At first, morphological preprocessing is applied for the preliminary selection of all possible candidate target regions. Morphological operations decompose the given input image into a filtered image. Clutter rejection, i.e. the classification between desired target and background, is done by means of Probabilistic neural network (PNN). For most cases, only the samples of the desired target images are used for the training purposes, which are not adequate for cases, where the target is almost blended with the background. For instance, target like objects may be present in the region of interest (ROI) and there is very low contrast difference between target and background. Horizontal and vertical convolution with wavelet low pass filter coefficients serves to extract features for training the PNN. In this paper, an improved clutter rejecter is presented to overcome the inferior classification performance of alternate techniques for poorly centered targets by moving the marked candidate target window in suitable directions with respect to the center of the potential target patch to extract ROIs from each detected target region. Results are shown for introductory detection-classification, and on improved performance of the clutter rejecter, by considering several shifted ROIs to accurately classify the true target from the clutter. Test results confirm the excellent performance of the detector and the clutter rejecter when both target and background features are used for training, and several shifted ROIs are considered for precise classification of each ROI marked by the detector.

The problem of target detection in a sequence of infrared images is not an easy task when the target is small, faint and obscured. The problem grows more complex when the target is embedded in a highly structured (correlated) background. In this paper, a new detector for small IR targets is proposed. This detector consists of three main components: a local whitening (demeaning) filter, an orthogonal image modeling algorithm, known as fast orthogonal search (FOS), and finally a first order statistical analysis is exploited for further reduction of false alarms. Experimental results of using the above detector to detect real infrared targets are also included. These results demonstrate that the new detector yields a promising solution for the detection problems of small targets.

Passive radar systems exploit illuminators of opportunity, such as TV and FM radio, to illuminate potential targets. Doing so allows them to operate covertly and inexpensively. Our research seeks to enhance passive radar systems by adding automatic target recognition (ATR) capabilities. In previous papers we proposed conducting ATR by comparing the radar cross section (RCS) of aircraft detected by a passive radar system to the precomputed RCS of aircraft in the target class. To effectively model the low-frequency setting, the comparison is made via a Rician likelihood model. Monte Carlo simulations indicate that the approach is viable. This paper builds on that work by developing a method for quickly assessing the potential performance of the ATR algorithm without using exhaustive Monte Carlo trials. This method exploits the relation between the probability of error in a binary hypothesis test under the Bayesian framework to the Chernoff information. Since the data are well-modeled as Rician, we begin by deriving a closed-form approximation for the Chernoff information between two Rician densities. This leads to an approximation for the probability of error in the classification algorithm that is a function of the number of available measurements. We conclude with an application that would be particularly cumbersome to accomplish via Monte Carlo trials, but that can be quickly addressed using the Chernoff information approach. This application evaluates the length of time that an aircraft must be tracked before the probability of error in the ATR algorithm drops below a desired threshold.

In this paper we report on the development of a technique for adaptive selection of polarization ellipse tilt, and ellipticity angles such that the targets separation from clutter is maximized. From the radar scattering matrix [S] and its complex components, in-phase and quadrature phase, the elements of Mueller/Kennaugh matrix are obtained. Then by means of polarization synthesis, the radar cross section of the radar scatterers at different transmitting and receiving polarization states are obtained. By designing a maximum average correlation height filter, a target vs. clutter distance measure is derived as function of four transmit and receive polarization state angles. The results of applying this method on real synthetic aperture radar imagery indicate a set of four transmit and receive angles which lead to maximum target vs. clutter discrimination. These optimum angles are different for different targets. Hence by adaptive control of the state of polarization of polarimetric radar, one can noticeably improve discrimination of targets from clutter.

A synthetic aperture radar (SAR) automatic target recognition (ATR) system based on the minimum noise and correlation energy (MINACE) distortion-invariant filter (DIF) is presented. A set of MINACE filters covering different aspect ranges is synthesized for each object using a training set of images of that object and a validation set of confuser and clutter images. No prior DIF work addressed confuser rejection. We also address use of fewer DIFs per object than prior work did. The selection of the MINACE filter parameter c for each filter is automated using training and validation sets. The system is evaluated using images from the Moving and Stationary Target Acquisition and Recognition (MSTAR) public database. The classification scores (PC) and the number of false alarm scores for confusers and clutter (PFA and PCFA respectively) are presented for the benchmark three-class MSTAR database with object variants and two confusers. The pose of the input test image is not assumed to be known, thus the problem addressed is more realistic than in prior work, since pose estimation of SAR objects has a large margin of error. Results for both confuser and clutter rejection are presented.

Ultra-wideband ground penetrating radar (GPR) systems are useful for extracting and displaying information for target recognition purposes. The frequency content of projected signals is designed to match the size and type of prospective targets and environments. Target signatures whether in the time, frequency, or joint time-frequency domains, will substantially depend on the target's burial conditions such as the type of soil, burial depth, and the soil’s moisture content. Such returned echoes from two targets for several moisture contents and burial depths in a soil with known electrical properties were simulated earlier by using a Method-of-Moments (MoM) code. The signature template of each target was computed using a time-frequency distribution of the returned echo when the target was buried at standard conditions, namely, a selected depth in the soil with a selected moisture content. For any returned echo the relative difference between the likewise computed target signature and a template signature was computed. That signature difference, chosen as objective function, or cost function, could then be minimized by adjusting the depth and moisture content, now taken to be unknown parameters. This can be done using the differential evolution method (DEM) together with our target translation method (TTM). The template that gave the smallest value of the minimized objective function for the returned echo signified the classification, and the corresponding values of the depth and moisture parameters turned out to be good predictions of the actual target depth and soil moisture content. Here, we implement a more efficient and faster running version of this classification method on a stepped-frequency continuous-wave (SFCW) GPR system. We demonstrate the ability to classify mines or mine-like targets buried underground from measured GPR signals. The targets are buried either in an indoor sandbox or in a test field at the Swedish Explosive Ordnance Disposal and Demining Center (SWEDEC) at Eksjo, Sweden.

With research on autonomous underwater vehicles for minehunting beginning to focus on cooperative and adaptive behaviours, some effort is being spent on developing automatic target recognition (ATR) algorithms that are able to operate with high reliability under a wide range of scenarios, particularly in areas of high clutter density, and without human supervision. Because of the great diversity of pattern recognition methods and continuously improving sensor technology, there is an acute requirement for objective performance measures that are independent of any particular sensor, algorithm or target definitions. This paper approaches the ATR problem from the point of view of information theory in an attempt to place bounds on the performance of target classification algorithms that are based on the acoustic shadow of proud targets. Performance is bounded by analysing the simplest of shape classification tasks, that of differentiating between a circular and square shadow, thus allowing us to isolate system design criteria and assess their effect on the overall probability of classification. The information that can be used for target recognition in sidescan sonar imagery is examined and common information theory relationships are used to derive properties of the ATR problem. Some common bounds with analytical solutions are also derived.

Sidescan sonar is increasingly accepted as the sensor of choice for sea minehunting over large areas in shallow water. Automatic Target Recognition (ATR) algorithms are therefore being developed to assist and, in the case of autonomous vehicles, even replace the human operator as the primary recognition agent deciding whether an object in the sonar imagery is a mine or simply benign seafloor clutter. Whether ATR aids or replaces a human operator, a natural benchmark for judging the quality of ATR is the unaided human performance when ATR is not used. The benchmark can help when estimating the performance benefit (or cost) of switching from human to automatic recognition for instance, or when planning how human and machine should best interact in cooperative search operations. This paper reports a human performance study using a large library of real sonar images collected for the development and testing of ATR algorithms. The library features 234 mine-like man-made objects deployed for the purpose, as well as 105 instances of naturally occurring clutter. The human benchmark in this case is the average of ten human subjects expressed in terms of a receiver operating characteristic (ROC) curve. An ATR algorithm for man-made/natural object discrimination is also tested and compared with the human benchmark . The implications of its relative performance for the integration of ATR are considered.

Growing military requirements and shorter timelines are placing greater demands on imagery analysts. At the same time, advances in sensor technology have vastly increased the quantity and types of imagery data available. Together, these factors are driving toward greater reliance on automated exploitation tools, such as automated target cueing (ATC). Several studies indicate that operational performance depends not only on the accuracy of the ATC algorithm, but also on effectively conveying the ATC information to the user. Sonification, the presentation of information through audio signals, provides a novel method for assisting analysts with visual search tasks. This paper presents a recent proof-of-concept experiment in which analysts search for geometric targets in synthetic, two-band color imagery. The performance results indicate that sonification can enhance performance, particularly through false alarm mitigation. The range of performance across users also suggests that user training may play a big role in effective operational use of sonification methods.

In this paper we describe an acoustic weapons detection concept that is based on ultrasonics and nonlinear acoustics. An ultrasonic projector is used to create an acoustic field at the site of inspection. The field is composed of multiple ultrasonic waves interacting at the interrogation site. The ultrasonic field creates acoustic interactions at that site which are used as the primary probe. The acoustic field is tailored to excite the target in an optimum fashion for weapons detection. In this presentation, we present aspects of this approach highlighting its ability to confine the interrogation field, create a narrow-band probing field, and the ability to scan that acoustic field to image objects. Ultrasonic propagation parameters that influence the field will be presented as will data of field characteristics. An image obtained with this system will be shown, demonstrating its capability to achieve high resolution. Effects of cloth over a weapon are shown to alter the image, yet not hide the weapon. Luna will report on its most recent findings as to the nature of this detection technology and its ability to generate information important to CWD.

In this paper we present an approach for signal enhancement of sonar signals. Work reported is based on sonar data collected by the Volume Search Sonar (VSS), as well as VSS synthetic data. The Volume Search Sonar is a beamformed multibeam sonar system with 27 fore and 27 aft beams, covering almost the entire water volume (from above horizontal, through vertical, back to above horizontal). The processing of a data set of measurement in shallow water is performed using the Fractional Fourier Transform algorithm. The proposed technique will allow efficient determination of seafloor bottom characteristics and bottom type using the reverberation signal. A study is carried out to compare the performance of the presented method with conventional methods. Results are shown and future work and recommendations are presented.

LADAR imagery provides the capability to represent high resolution detail of 3D surface geometry of complex targets. In previous work we exploited this capability for automatic target recognition (ATR) by developing matching algorithms for performing surface matching of 3D LADAR point clouds with highly-detailed target CAD models. A central challenge in evaluating ATR performance is characterizing the degree of problem difficulty. One of the most important factors is the inherent similarity of target signatures. We've developed a flexible approach to target taxonomy based on 3D shape which includes a classification framework for defining the target recognition problem and evaluating ATR algorithm performance. The target model taxonomy consists of a hierarchical, tree-structured target classification scheme in which different levels of the tree correspond to different degrees of target classification difficulty. Each node in the tree corresponds to a collection of target models forming a target category. Target categories near the tree root represent large and very general target classes, exhibiting large interclass distance. Targets in these categories are easily separated. Target categories near the tree bottom represent very specific target classes with small interclass distance. These targets are difficult to separate. In this paper we focus on creation of optimal categories. We develop approaches for optimal aggregation of target model types into categories which provide for improved classification performance. We generate numerical results using match scores derived from matching highly-detailed CAD models of civilian ground vehicles.

Automatic target classification in general is complicated owing to the influence of pose, articulation, and overall viewing geometry on two dimensional SAR data. Three dimensional (3D) data, however, affords the opportunity to develop robust classification techniques independent of those issues. Based on geometric invariants, discriminants can be obtained assuming the target or its phase center lattice can be well modelled by 3D geometries subject to independent rigid body motions, (i.e. reflection, rotation, and translation). Toward this end, we present recent results in the development of a unique 3D classification algorithm. The concepts herein are developed for the full 3D observation space. In particular, we discuss several discrimination metrics based on a target's geometry. As such, they are necessarily invariant to pose and articulation and consequently provide robust classification performance. These geometric-invariant discriminants are concisely expressed as equations unique to a single target structure, or to the spatial interrelationships of multiple structures (this addresses the articulation problem). Once established, these equations can subsequently be used to properly classify the structure or structures at a later time without the need for explicit knowledge of the 3D orientation of the structures within the field of view. We present the mathematical basis behind these classification schemes, discuss implementation concepts, and finish by demonstrating these techniques on synthetic data.

Automatic Target Recognition (ATR) algorithms are extremely sensitive to differences between the operating conditions under which they are trained and the extended operating conditions in which the fielded algorithms operate. For ATR algorithms to robustly recognize targets while retaining low false alarm rates, they must be able to identify the conditions under which they are operating and tune their parameters on the fly. In this paper, we present a method for tuning the parameters of a model based ATR algorithm using estimates of the current operating conditions. The problem has two components: 1) identifying the current operating conditions and 2) using that information to tune parameters to improve performance. In this paper, we explore the use of a reinforcement learning technique called tile coding for parameter adaptation. In tile coding, we first define a set of valid states describing the world (the operating conditions of interest, such as the level of obscuration). Next, actions (or parameter settings used by the ATR) are defined that are applied when in that state. Parameter settings for each operating condition are learned using an off-line reinforcement learning feedback loop. The result is a lookup table to select the optimal parameter settings for each operation condition. We present results on real LADAR imagery based on parameter tuning learned off-line using synthetic imagery.

Spin images originated within the robotics group at Carnegie Mellon University and are representations of 3-space surface regions. This representation provides a means for surface matching that is invariant to rigid body rotations and translations while being robust in the presence of 3D image noise, clutter, and surface occlusion. Of particular interest is the viability of using spin images to differentiate between two object classes in 3D imagery where there is significant intra-class diversity, e.g. to differentiate between wheeled and tracked vehicles. The specificity of spin map representations in differentiation of wheeled and tracked vehicles is statistically characterized. Using synthetic imagery of various wheeled and tracked vehicles, the class separability of wheeled vs. tracked vehicle spin image sets is nonparametrically quantified via entropic characterization as well as the Friedman-Rafsky two-sample test statistic. Additionally, class separability is analyzed in lower dimensional feature spaces generated via the Hotelling transform as well as a random projection method, comparing and contrasting the spin map class differentiation in the original and transformed data sets.

Vibration measurement using coherent laser radar (LADAR) is a promising way to identify air targets at long range. Laser vibrometers can remotely measure the velocity of micrometric displacements and thus exhibit the target surface vibration frequencies. Some of these frequencies are modal frequencies, which result from the target structure. They define a unique signature and allow target identification to be performed. As vibration amplitudes are not reliable, we choose to consider only frequency positions. In this article, we explain an "extended identification" method which takes into account cumulative signatures in space and time to improve global system identification performance. Using a nearest neighbor classifier and a suitable metric taking into account a simple off-line processing of measured data, the recognition algorithm leads to good identification rates and very low rejection rates for a nine class problem. We show a strong improvement of the identification rate thanks to the "extended identification" method.

The paper compares the target identification performance of conventional model matching criteria and of new probabilistic techniques based on Bayesian hypothesis generation and verification. Match techniques are categorized into two types: those requiring target segmentation results and those which do not. Applied to low-resolution laser radar images of military vehicles, deterministic techniques using no segmentation results had the lowest target identification rates. New probabilistic techniques using no segmentation results are introduced, having significantly higher target identification rates than the best known deterministic procedures. The best results were attained by a probabilistic matching approach requiring target segmentation. Using certain simplifying assumptions, the latter technique can be reformulated as a deterministic procedure, involving no probabilities on scene parameters, and having almost the same target identification performance.

The problem of constructing a single classifier when multiple phenomenologies are measured by different sensor types is made more difficult because features take diversified forms, and classifiers built from them have variable performance. For example, features can be continuous or binary valued (as in discrete labels), or be composed of incompatible structural primitives. Therefore, it is difficult to lump all of these features together into a single classifier for decision making. This realization leads to the combined use of multiple classifiers. The solution presented in this paper describes the formulation and development of: A computational procedure for computing approximate hyperplane decision boundaries to achieve a balanced classifier. Achieving a minimum Bayes-risk balanced classifier as a convex combination of balanced classifiers. This is done for both independent and correlated cases. Convex combinations of balanced classifiers are balanced. However, our research has further generalized this concept by computing optimal convex combinations of classifiers so as to also attain the property of being minimum Bayes-risk for the combined classifier. The principle exploited was to incorporate either the decisions or the decision statistics of the individual classifiers within a combined confusion matrix considering both the correlated and independent cases. This was posed as an optimization problem to be approached via Markov-Chain Monte Carlo methods. Some preliminary results are shown.

The detection of subpixel targets in hyperspectral images is complicated by interference arising from other background materials. This paper describes three target detection algorithms implemented in Data Fusion Corporation's HYPERTOOLS, a suite of hyperspectral image analysis tools. The matched subspace filter (MSF) is a generalized likelihood ratio test designed to detect target signatures while suppressing known interference signatures in a hyperspectral image. The fill-factor matched subspace filter (FFMSF) and the mixture-modeled matched subspace filter (MMMSF) extend the MSF by fusing geometrical (i.e., material abundance) and statistical (i.e., an assessment of the applicability of a linear replacement mixture model) information with the MSF output. The MSF, FFMSF, and MMMSF require one, two, and three thresholds, respectively. Automated means of determining these thresholds are proposed and justified. The MSF is further designed to allow the processing of multirank target and interference spectral matrices. As more information about a target or targets is included in the MSF, the detection performance of the MSF is expected to improve. If the target and/or interference matrices are singular or nearly singular, however, the performance of the MSF may instead be degraded. Singular value decomposition (SVD) may be employed to prepare spectral data matrices for optimal performance of the MSF. Although the use of singular value decomposition for preprocessing data matrices is well-known in signal processing, the determination of thresholds for the selection of left-singular vectors spanning the data space remains more of “an art.” An automated method for determining the number of useful left-singular vectors is proposed based on an interpretation of the singular values and on the analysis of the dimensions of the measurement space.

Human vision derives and uses an entirely different type and quantity of visual information than physical vision sensors. Autonomous visual systems that seek to emulate the outstanding capability of human vision should also use this very different information. It is shown how this information can be applied in a micro UAV to sense range for collision avoidance.

The majority of automatic target recognition (ATR) studies are formulated as a traditional classification problem. Specifically, using a training set of target exemplars, a classifier is developed for application to isolated measurements of targets. Performance is assessed using a test set of target exemplars. Unfortunately, this is a simplification of the ATR problem. Often, the operating conditions differ from those prevailing at the time of training data collection, which can have severe effects on the obtained performance. It is therefore becoming increasingly recognised that development of robust ATR systems requires more than just consideration of the traditional classification problem. In particular, one should make use of any extra information or data that is available. The example in this paper focuses on a hybrid ATR system being designed to utilise both measurements from identity sensors (such as radar profiles) and motion information from tracking sensors to classify targets. The first-stage of the system uses mixture-model classifiers to classify targets into generic classes based upon data from (long range) tracking sensors. Where the generic classes are related to platform types (e.g. fast-jets, heavy bombers and commercial aircraft), the initial classifications can be used to assist the military commander's early decision making. The second-stage of the system uses measurements from (closer-range) identity sensors to classify the targets into individual target types, while taking into account the (uncertain) outputs from the first-stage. A Bayesian classifier is proposed for the second-stage, so that the first-stage outputs can be incorporated into the second-stage prior class probabilities.

Feature extraction and selection are important problems in statistical learning. We study the relationships between two previously proposed principles for their optimal solution: the minimization of Bayes error and the maximization of mutual information between features and class labels. It is shown that a quantity which provides insight on this relationship is the set of non-increasing probability mass functions (NIPMFs). We derive some basic properties of the members of this set, show that any classification problem defines an ensemble of NIPMFs, and that the probability distribution of this ensemble uniquely determines the associated Bayes error and mutual information. These results are then used to show that, when the classification problem is binary and some generic constraints hold, the optimal feature space is the same under the two formulations.

Robust real-time recognition of multiple targets with varying pose requires heavy computational loads, which are often too demanding to be performed online at the sensor location. Thus an important problem is the performance of ATR algorithms on highly-compressed video sequences transmitted to a remote facility. We investigate the effects of H.264 video compression on correlation-based recognition algorithms. Our primary test bed is a collection of fifty video sequences consisting of long-wave infrared (LWIR) and mid-wave infrared (MWIR) imagery of ground targets. The targets are viewed from an aerial vehicle approaching the target, which introduces large amounts of scale distortion across a single sequence. Each sequence is stored at seven different levels of compression, including the uncompressed version. We employ two different types of correlation filters to perform frame-by-frame target recognition: optimal tradeoff synthetic discriminant function (OTSDF) filters and a new scale-tolerant filter called fractional power Mellin radial harmonic (FPMRH) filters. In addition, we apply the Fisher metric to compressed target images to evaluate target class separability and to estimate recognition performance as a function of video compression rate. Targets are centered and cropped according to ground truth data prior to separability analysis. We compare our separability estimates with the actual recognition rates achieved by the best correlation filter for each sequence. Numerical results are provided for several target recognition examples.

A major challenge for ATR evaluation is developing an accurate image truth that can be compared to an ATR algo-rithm's decisions to assess performance. While many standard truthing methods and scoring metrics exist for stationary targets in still imagery, techniques for dealing with motion imagery and moving targets are not as prevalent. This is par-tially due to the fact that the moving imagery / moving targets scenario introduces the data association problem of as-signing targets to tracks. Video datasets typically contain far more imagery than static collections, increasing the size of the truthing task. Specifying the types and locations of the targets present for a large number of images is tedious, time consuming, and error prone. In this paper, we present an updated version of a complete truthing system we call the Scoring, Truthing, And Registration Toolkit (START). The application consists of two components: a truthing compo-nents that assists in the automated construction of image truth, and a scoring component that assesses the performance of a given algorithm relative to the specified truth. In motion imagery, both stationary and moving targets can be de-tected and tracked over portions of a motion imagery clip. We summarize the capabilities of START with emphasis on the target tracking and truthing diagnostics. The user manually truths certain key frames, truth for intermediate frames is then inferred and sets of diagnostics verify the quality of the truth. If ambiguous situations are encountered in the inter-mediate frames, diagnostics flag the problem so that the user can intervene manually. This approach can dramatically reduce the effort required for truthing video data, while maintaining high fidelity in the truth data. We present the results of two user evaluations of START, one addressing the accuracy and the other focusing on the human factors aspects of the design.

Correlation filters are attractive for automatic target recognition (ATR) applications due to such attributes as shift invariance, distortion tolerance and graceful degradation. Composite correlation filters are designed to handle target distortions by training on a set of images that represent the expected distortions during testing. However, if the distortion can be described algebraically, as in the case of in-plane rotation and scale, then only one training image is necessary. A recently introduced scale-tolerant correlation filter design, called the Minimum Average Correlation Energy Mellin Radial Harmonic (MACE-MRH) filter, exploits this algebraic property and allows the user to specify the scale response of the filter. These filters also minimize the average correlation energy in order to help control the sidelobes in the correlation output and produce sharper, more detectable peaks. In this paper we show that applying non-linearities in the frequency domain (leading to fractional power scale-tolerant correlation filters) can significantly improve the resulting peak sharpness, yielding larger peak-to-correlation energy values for true-class targets at various scales in a scene image. We investigate the effects of fractional power transformations on MACE-MRH filter performance by using a testbed of fifty video sequences consisting of long-wave infrared (LWIR) imagery, in which the observer moves along a flight path toward one or more ground targets of interest. Targets in the test sequences suffer large amounts of scale distortion due to the approach trajectory of the camera. MACE-MRH filter banks are trained on single targets and applied to each sequence on a frame-by-frame basis to perform target detection and recognition. Recognition results from both fractional power MACE-MRH filters and regular MACE-MRH filters are provided, showing the improvement in scale-tolerant recognition from applying fractional power non-linearities to these filters.

Video compression is critical to many real world applications where the data is being transmitted over data links with limited bandwidth. Previously, information metrics have been used to assess the distortion of target signatures with differing degrees of compression. A recent update to the emerging H.264 standard, called Fidelity Range Extensions (FRExt), permits applications up to 12-bits/sample, and even monochrome only, which are especially relevant for IR applications. This paper tests this new FRExt technology on actual IR sensor data to obtain preliminary results on the compression of 12-bit IR data using state of the art video compression technology.

In this paper, we proposed a new architecture, called nonzero-order fringe-adjusted joint transform correlator (FJTC) using a binary random phase mask, for real time pattern recognition applications. The binary random phase mask encodes the reference image by two equal probability phases before it is introduced in the joint input image. The joint power spectrum is then multiplied by the same phase mask to remove the zero-order term and false alarms that may be generated in the correlation plane due to the presence of multiple identical target or non-target objects in the input scene. The criteria used for measuring the performance of nonzero-order fringe-adjusted JTC include correlation peak intensity, peak-to-correlation energy, and peak-to-sidelobe ratio. Detailed analysis for the proposed nonzero-order FJTC using binary random phase mask is presented. Simulation results verify the effectiveness of the proposed technique.

In ATR applications, each feature is a convolution of an image with a filter. It is important to use most discriminant features to produce compact representations. We propose two novel subspace methods for dimension reduction to address limitations associated with Fukunaga-Koontz Transform (FKT). The first method, Scatter-FKT, assumes that target is more homogeneous, while clutter can be anything other than target and anywhere. Thus, instead of estimating a clutter covariance matrix, Scatter-FKT computes a clutter scatter matrix that measures the spread of clutter from the target mean. We choose dimensions along which the difference in variation between target and clutter is most pronounced. When the target follows a Gaussian distribution, Scatter-FKT can be viewed as a generalization of FKT. The second method, Optimal Bayesian Subspace, is derived from the optimal Bayesian classifier. It selects dimensions such that the minimum Bayes error rate can be achieved. When both target and clutter follow Gaussian distributions, OBS computes optimal subspace representations. We compare our methods against FKT using character image as well as IR data.

Nearest neighbor classifiers are one of most common techniques for classification and ATR applications. Hastie and Tibshirani propose a discriminant adaptive nearest neighbor (DANN) rule for computing a distance metric locally so that posterior probabilities tend to be homogeneous in the modified neighborhoods. The idea is to enlongate or constrict the neighborhood along the direction that is parallel or perpendicular to the decision boundary between two classes. DANN morphs a neighborhood in a linear fashion. In this paper, we extend it to the nonlinear case using the kernel trick. We demonstrate the efficacy of our kernel DANN in the context of ATR applications using a number of data sets.

Many classifiers have been proposed for ATR applications. Given a set of training data, a classifier is built from the labeled training data, and then applied to predict the label of a new test point. If there is enough training data, and the test points are drawn from the same distribution (i.i.d.) as training data, then many classifiers perform quite well. However, in reality, there will never be enough training data or with limited computational resources we can only use part of the training data. Likewise, the distribution of new test points might be different from that of the training data, whereby the training data is not representative of the test data. In this paper, we empirically compare several classifiers, namely support vector machines, regularized least squares classifiers, C4.4, C4.5, random decision trees, bagged C4.4, and bagged C4.5 on IR imagery. We reduce the training data by half (less representative of the test data) each time and evaluate the resulting classifiers on the test data. This allows us to assess the robustness of classifiers against a varying knowledge base. A robust classifier is the one whose accuracy is the least sensitive to changes in the training data. Our results show that ensemble methods (random decision trees, bagged C4.4 and bagged C4.5) outlast single classifiers as the training data size decreases.

While modern sensors allow vast amounts of data to be collected in seconds, it can take weeks or months to analyze the data and determine sensor performance. Because of this lag between data collection and usable results, issues such as sensor calibration and algorithm biases are not detected until well after an experiment or test, when it is too late to correct them. Recognizing that a rapid performance snapshot would be extremely valuable in many situations, the AFRL COMPASE Center developed tools to produce receiver operating characteristic (ROC) curves in near real-time for an advanced technology demonstration (ATD) program. These tools, called real-time (RT-ROC) and identification (ID-ROC), gave the evaluation team timely insight into overall system performance, and when substandard results were obtained, diagnostic tests were initiated to determine the underlying causes. These tools have been experiment demonstrated, allowing the COMPASE team to find and fix a sensor error in a matter of hours. This paper will concentrate on RT-ROC and will address analysis tool requirements, operation of the tool during experiments, a walkthrough of the tool using simulated data, and future uses for this application.

This investigation discusses the challenge of target classification in terms of intrinsic dimensionality estimation and selection of appropriate feature manifolds with object-specific classifier optimization. The feature selection process will be developed via nonlinear characterization and extraction of the target-conditional manifolds derived from the training data. We investigate defining the feature space used for classification as a class-conditioned nonlinear embedding, i.e., each training and test image is embedded in a target-specific embedding and the resultant embeddings are used for statistical characterization. We compare and contrast this novel embedding technique with Principal Component Analysis. The α-Jensen Entropy Difference measure is used to quantify the object-conditioned separation between the target distributions in the feature spaces. We discuss and demonstrate the effect of feature space extraction on classification efficacy.

A new generation of target recognition systems must be based on the principles of image understanding and active vision. The implementation of both principles is possible in the form of Network-Symbolic systems. Instead of precise computations of 3-dimensional models a network-symbolic system converts image information into an "understandable" Network-Symbolic format, which is similar to relational knowledge models. An Image/Video Analysis that is based on Network-Symbolic models differs from the traditional linear bottom-up "segmentation-grouping-learning-recognition" approach. It is a combination of recursive hierarchical bottom-up and top-down processes. Logic of visual scenes can be captured in the Network-Symbolic models and used for the disambiguation of visual information, including target detection and identification. View-based object recognition is a hard problem for traditional algorithms that directly match a primary view of an object to a model. In Network-Symbolic Models, the derived structure and not the primary view is a subject for recognition. Such recognition is not affected by local changes and appearances of the object from a set of similar views. Network-Symbolic systems can be treated as a new type of Multi-Agent systems that can better interpret visual information for automatic target detection and identification required by a new generation of smart defense systems.

If the binary image pattern, e.g., the edge-detected boundary of an object, is varying in real time among several extreme boundaries, then learning just the extreme boundaries by an OLNN (one-layered neural network) will allow the OLNN to recognize any unlearned, time-varying patterns of the object varying among these extreme boundaries. This is possible because of the unique property of CONVEX LEARNING existing in the OLNN. This paper will first derive this property from mathematical point of view, and then verify it with some simple experiments. The main advantage of this neural network is that it can recognize very similar objects not only from the static patterns it learns but also from the ways how these objects vary in real time even these varying patterns are NOT learned one by one at each time. Consequently the recognition is much more accurate than just learning the static patterns alone.

When the M input training class patterns are represented by N-dimension analog vectors U_m, m=1 to M, the output P-bit binary classification vectors V_m, M=1 to M, of a one-layered, feed-forward neural network (OLNN) can be represented geometrically by P dichotomizing hyper-planes going through the origin of the N-dimension Euclidean coordinate in the N-space. In general, all these P planes divide the N-space into 2^P hyper-cones. Each cone contains one U_m and each cone corresponds to one V_m. Learning of the OLNN is then equivalent to establishing these P planes geometrically in the N-space such that, after the learning, if a test pattern vector T, not necessarily equal to any class pattern U_m, falls into the m-th cone (m=1 to 2^P) established by these P planes, this T will also be recognized as V_m. The robustness of this recognition is seen now to be equivalent to the geometrically allowed variation range of U_m in the m-th cone. This allowable range can be systematically adjusted for each cone during the learning process. This paper reports the optimum method of adjusting these variation ranges such that any unknown T containing environmental noise not included in the training can still be recognized with maximum accuracy.

This paper proposes a novel algorithm for automatically and quickly extracting the external edge of the moving object. The algorithm uses both spatial and temporal information, which is abundant in image sequences and has not usually been efficiently used in conventional algorithm. Especially most attention of this algorithm is paid for the moving object external-edges. This orientation is designed to deal with the overlap of moving objects in successive frames. After the temporal segmentation the external edge of moving object is detected. Differing from current algorithms the fusing way of the temporal and spatial information has been progressed. Here the algorithm only uses spatial information of the temporal external edge to refine the edge points for getting the accurate boundary. This method will greatly decrease the compute and achieve the fast process in the real-time system. Though the simulation, this algorithm can get perfect result.

The aim of investigation consists in development of a formal image representation, in whose framework the most relevant information can be extracted from images. Constructing the models of images is considered as a task of inductive inference. The conventional criterions for choosing the best model are based on the Bayesian rule. However there is one classical problem of defining the a priori probabilities of models. The generally adopted approach for overcoming this difficulty is to use the Minimum Description Length (MDL) principle. In the task of interpretation of visual scenes the a priori probabilities of realizations of images are assigned by their representation language. In our work we study the hierarchical structural descriptions of images. A problem of selection of alphabet of structural elements is addressed. Such the commonly used structural elements as the straight lines, angles, arcs, and others are considered, and their usage is grounded on the base of the amount of information contained in them. The composite structural elements can be formed within the framework of hierarchical representations. The grouping rules are generally based on some similarities in the elements. Hence the descriptions of these elements contain the positive mutual information. Such the approach permits to proof the usage of these structural elements, to choose rationally their types, and to elaborate a rigorous criterion of grouping. The results of research implemented in the form of computer programs showed the appropriateness of this approach.

The investigation presented in this article continues our long-term efforts directed towards the automatic structural matching of aerospace photographs. An efficient target independent hierarchical structural matching tool was described in our previous paper, which, however, was aimed mostly for the analysis of 2D scenes. It applied the same geometric transformation model to the whole area of image, thus it was nice for the space photographs taken from rather high orbits, but it often failed in the cases when the sensors were positioned near the 3D scenes being observed. Different transformation models should be applied to different parts of images in the last case, and finding a correct separation of image into the areas of homogeneous geometric transformations was the main problem. Now we succeeded in separating the images of scenes into the surfaces of different objects on the base of their textural and spectral features, thus we have got a possibility of separate matching the sub-images corresponding to such objects applying different transformation model to each such sub-image. Some additional limitations were used in the course of such separation and matching. In particular, the a priory assumptions were applied in different cases about the possible geometry of scenes, rules of illumination and shadowing, thus the aerospace photographs, indoor scenes, or images of aircrafts were analyzed in slightly differing ways. However the additional limitations applied could be considered as very general and are worth to be used in a wide sphere of practical tasks. The automatic image analysis was successful in all considered practical cases.

A multi-resolution approach to automatic target recognition is described that employs a hybrid evolutionary algorithm (HEA) and image transform in a form of image local response. Given images of the targeted area (TA) and the targeted object (TO) located in TA, the proposed method repeatedly applies cross-correlation on different resolution levels (zooming in), in order to find the area TA and the object TO in the large-scale image of the region of interest (ROI). Both images of ROI and TA undergo peculiar transformation called image local response. Given geometric transformation T(V) of the images under specified parameter vector V, image local response is defined as an image transform R(V) that maps an image into itself, with the small perturbation of the parameter vector V. Unit variations of the components of the parameter vector V are applied to the image, and the corresponding variations of the least squared difference of the gray levels of the two images (i.e., before and after the parameter variation) form an image response matrix M(V). Cross-correlation of the response matrices built for ROI and TA outlines a potential range of resolutions of the TA. A hybrid Evolutionary algorithm can be applied then, in order to find the correct parameters V for TA with the reference to ROI.

We present analysis and some experiments for rapid recognition and detection of hidden objects (i.e. human figures) using terahertz radiation. T-rays have a unique advantage, namely high reflectivity compared with all other electromagnetic waves and the ability to pass through most building materials. Further, the high frequency range of the terahertz band has the potential of reduced equipment size as compared with current radar technology. Imaging, at these frequencies, is developing fairly rapidly as compared with communication. However, usual imaging in the form of SAR, ISAR or electro-optical imaging takes a long time due to the large dwell time to acquire a single image. In this report, we investigate techniques for rapid classification using one dimension high resolution range profiles. Methods of statistical pattern recognition will be applied for identification of the object.