The feasibility of ultra-wideband (UWB) radar systems for extracting signature information useful for target recognition purposes has been demonstrated previously. An UWB radar system emits either an extremely short pulse, impulse, or a frequency modulated signal (e.g., sweep or step-frequency). The frequency content of the emitted signals is designed to match the size and kind of typical targets and environments. We study the performance of two different radar systems, each intended for use as a ground penetrating radar (GPR). These systems are an impulse radar and a stepped-frequency continuous-wave radar. The frequency content of the emitted pulses is about 300 - 3,000 MHz with both systems. The same antenna system consisting of two crossed dipoles (transmitting and receiving) is used with both radar systems and it is directed upward in an outdoor setup. The targets are metal spheres of two sizes and each of them is placed above the antenna system on a thick foam plate made of polystyrene, which is virtually transparent for microwaves. Backscattered echoes from each target are analyzed in the frequency domain using traditional Fourier transforms and in the joint time- frequency domain using Wigner distributions (WD). The measured and digitally processed waveforms are then compared with theoretical predictions and the relative merits of the two UWB radar systems are evaluated. The results of our investigation serve to assess the best choice of GPR system for extracting discriminating target signature information.

In an effort to improve the usefulness of computer classifiers for military applications, the U.S. Army Research Laboratory has begun to develop a database of synthetic infrared target chips. Once created, this database will aid in the training and testing of both human and computer classifiers, and will provide a way to train classifiers on targets and clutter environments with little real data available. Results presented below will show that classifier performance trained on synthetic data is improving but is, in general, poorer than when trained on real data, that individual synthetic target models perform much better than other models, providing evidence that better overall performance may yet be achievable, that synthetic data thus far created is highly self-similar and/or to some unknown extent represents real data not included in our database, and that enhanced performance of classifiers trained on small amounts of real data can be achieved by adding limited amounts of synthetic data.

To support Autonomous Target Acquisition (ATA) evaluation and trades analysis, the Air Force Research Laboratory, Advanced Guidance Division (AFRL/MNG) located at Eglin AFB has incorporated a general-purpose performance evaluation system into its Modular Algorithm Concept Evaluation Tool (MACET). The MACET performance evaluation system may be used for active, passive, or multi-sensor ATA analysis. It consists of two main elements: a relational, multi-user database engine and a database client application, the Performance Evaluation Tool (PET). The database engine serves a set of databases that are used to capture, catalog, and archive test results for various algorithms under varying condition and environments. The MACET PET client application is a data mining tool for exploring the ATA test results in the databases, computing standard ATA detection and classification performance metrics (e.g., detection probability, detection reliability, false alarm rate, probability of correct classification, confusion matrices) on user defined subsets of data, calculating test case parameter statistics, and generating performance comparison plots.

In matching of the images that are obtained from different sources one often observes inherent discrepancies that exist among them. These discrepancies can cause degradation in the object recognition performance, or errors in image registration system that are the basis of some automatic navigational systems. In this paper, a lower bound on the matching error is derived by using a model that treats the two images as being the two ends of a communication channel. Then rate distortion theory is applied to obtain a relationship between the matching error and the statistical properties of the images. The resulting bound can be used to predict the matching error, or used in the selection of the proper model and reference images, and in the performance evaluation of the target recognition systems.

Radar targets often have both specular and diffuse scatterers. A conditionally Rician model for the amplitudes of pixels in Synthetic Aperture Radar (SAR) images quantitatively accounts for both types of scatterers. Conditionally Rician models generalize conditionally Gaussian models by including means with uniformly distributed phases in the complex imagery. Qualitatively, the values of the two parameters in the Rician model bring out different aspects of the images. For automatic target recognition (ATR), log-likelihoods are computed using parameters estimated from training data. Using MSTAR data, the resulting performance for a number of four class ATR problems representing both standard and extended operating conditions is studied and compared to the performance of corresponding conditionally Gaussian models. Performance is measured quantitatively using the Hilbert-Schmidt squared error for orientation estimation and the probability of error for recognition. For the MSTAR dataset used, the results indicate that algorithms based on conditionally Rician and conditionally Gaussian models yield similar results when a rich set of training data is available, but the performance under the Rician model suffers with smaller training sets. Due to the smaller number of distribution parameters, the conditionally Gaussian approach is able to yield a better performance for any fixed complexity.

In target recognition in uncontrolled environments the test target may not belong to the prestored targets or target classes. Hence, in such environments the use of a typical classifier which finds the closest class still leaves open the question of whether the test target truly belongs to that class. To decide whether a test target matches a stored target, common approaches calculate a degree of similarity between the two targets using a similarity measure such as Euclidean distance, and make a decision based on whether the distance exceeds a (prespecified) threshold. Based on psychophysical studies, this is very different from, and far inferior to, human capabilities. In this paper we show a new approach where a neural network learns a decision boundary between the confirmation vs. rejection of a match with the help of a human critic. The decision boundary is a multidimensional surface, and models the human similarity measure for the recognition task at hand, thus avoiding metric similarity measures and thresholds. A case study in automatic aircraft recognition is shown. In the absence of sufficient real data, the approach allows us to specifically generate an arbitrarily large number of training exemplars projecting near the classification boundary. The performance of the trained network was comparable to that of a human expert, and far better than a network trained only on the available real data. Furthermore, the result were considerably better than those obtained using a Euclidean discriminator.

Many different automatic target recognition (ATR) approaches have had their performance quantified for over twenty years typically by plotting a receiver operating curve (ROC) of probability of detection and/or recognition versus some measure of false alarm or false alarm rate. These ROCs have been generated on static sets of test and training data. This data, in some cases, has had significantly varying levels of difficulty, however, the quantification of the data set difficulty has typically only been coarsely partitioned based on the time of day, the target operational state, the meteorological environment, and sometimes the terrain or location. In addition, there has been no generally useful comparative measure of the target signature knowledge base provided for ATR system 'training' versus the signatures of the same targets in the data used for test. In this paper, we illustrate the quantification of two 'information content' data metrics with an associated ATR performance. The first metric is a signal to clutter measure (SC), and the second is a knowledge base signature distortion measure (KBSD) of the 'closest' target training signature versus the target test signature. These metrics provide a new basis for truly objective ATR performance comparison.

The grand tour is a method for visualizing high dimensional data by presenting the user with a set of projections and the projected data. This idea was extended to multispectral images by viewing each pixel as a multidimensional value, and viewing the projections of the grand tour as an image. The user then looks for projections which provide a useful interpretation of the image, for example, separating targets from clutter. We discuss a modification of this which allows the user to select convolution kernels which provide useful discriminant ability, both in an unsupervised manner as in the image grand tour, or in a supervised manner using training data. This approach is extended to other window-based features. For example, one can define a generalization of the median filter as a linear combination of the order statistics within a window. Thus the median filter is that projection containing zeros everywhere except for the middle value, which contains a one. Using the convolution grand tour one can select projections on these order statistics to obtain new nonlinear filters.

This novel approach uses automatic target detection together with compression techniques to achieve intelligent compression by exploiting knowledge of the image content. Two techniques have been experimented with one using horizontal-vertical (HV) partitioned quadtrees the other a variant of entropy called approximate entropy. The object masks that are generated using either of the techniques (or indeed other feature detectors) effectively cue potential areas of interest for subsequent encoding using two 'intelligent' image compression techniques. In the first approach, lossless compression algorithms can be applied to regions of interest within the images so that their statistical properties can be preserved to allow detailed analysis or further processing while the remainder of the image can be compressed with lossy algorithms. The degree of lossy compression is dependent both on the information content as well as the bandwidth requirement. In the second approach a wavelet-based decomposition is applied in which selective destruction of wavelet coefficients is performed outside the cued areas of interest (in effect concentrating the wavelets in required areas) prior to the encoding with a version of the progressive SPIHT encoder. Results will illustrate how both these approaches can be used for the detection and compression of airborne reconnaissance imagery.

This paper examines the use of three feature sets for object classification in aerial imagery. The first feature set is based on affine invariant functions of the central moments computed on the objects within the image. The second feature set employed Zernike moment invariants and the third feature set utilized affine invariant functions of the central moments that are computed over a spline fit to the object boundary. The initial object locations were obtained using either a region of interest identification process based on low-level image processing techniques or a hand extraction process. A single nearest neighbor, k-nearest neighbors, and a weighted k-nearest neighbors classifier were employed to evaluate the utility of the various feature sets for both the hand extracted and region of interest identified objects. The performance of the full system is characterized via probability of detection and probability of false alarm.

The area of automatic target classification has been a difficult problem for many years. Many approaches involve extracting information from imagery through a variety of statistical filtering and sampling techniques, resulting in a reduced dimension feature vector that is the input for a learning algorithm. The Support Vector Machine (SVM) algorithm is a wide margin classifier that provides reasonable results for sparse data sets. This can allow one to avoid the feature extraction step and process images directly. The SVM algorithm has the additional benefits that there are few parameters to adjust and the solutions are unique for a given training set. We applied SVM to a variety of data sets, including character recognition, military vehicles and Synthetic Aperture Radar data, and compared the results to some standard neural network architectures. It was found that the SVM algorithm gave equivalent or higher correct classification results compared to the neural networks with some noted advantages.

Modeling of a target starts with a subject matter expert (SME) analysis of the available sensor(s) data. The SME then forms relationships between the data and known target attributes, called evidence, to support modeling of different types of targets or target activity. Speeds in the interval 10 to 30 knots and ranges less than 30 nautical miles are two samples of target evidence derived from sensor data. Evidence is then organized into sets to define the activities of a target and/or to distinguish different types of targets. For example, near an airport, target activities of takeoff, landing, and holding need to be evaluated in addition to target classification of civilian or commercial aircraft. This paper discusses a method for evaluation of the inferred activities over the finite evidence domain formed from the collection of models under consideration. The methodology accounts for repeated use of evidence in different models. For example, 'near an airport' is a required piece of evidence used repeatedly in the takeoff, landing, and holding models of a wide area sensor. Properties of the activity model evaluator methodology are discussed in terms of model construction and informal results are presented in a Boolean evidence type of problem domain.

A method for the automatic detection of tanks and other vehicles in infrared imagery will be described. First regions of interest in the infrared imagery are identified using a novel method that combines histogram specification, applying a fixed grayscale threshold to the image, and performing image labeling on the thresholded image. Features are next extracted from identified regions of interest. The features are input to a fuzzy inference system. The output of the fuzzy inference system is a target confidence value that is used to classify targets at objects of interest or clutter.

Our purpose is, in medium term, to detect in air images, characteristic shapes and objects such as airports, industrial plants, planes, tanks, trucks, ... with great accuracy and low rate of mistakes. However, we also want to value whether the link between neural networks and multi-agents systems is relevant and effective. If it appears to be really effective, we hope to use this kind of technology in other fields. That would be an easy and convenient way to depict and to use the agents' knowledge which is distributed and fragmented. After a first phase of preliminary tests to know if agents are able to give relevant information to a neural network, we verify that only a few agents running on an image are enough to inform the network and let it generalize the agents' distributed and fragmented knowledge. In a second phase, we developed a distributed architecture allowing several multi- agents systems running at the same time on different computers with different images. All those agents send information to a 'multi neural networks system' whose job is to identify the shapes detected by the agents. The name we gave to our project is Jarod.

This paper focuses on utilizing the associative capabilities of the Hopfield neural net in processing digitized speech and recovering erroneous speech segments and reconstructing noisy speech. The scope of this study is limited and the tests conducted are exploratory in nature. However, with a limited vocabulary that fits many practical applications, this study shows that digitized speech can be enhanced using properly trained recurrent networks such as the Hopfield neural net. The results indicate that a Hopfield neural network with sufficient associative memory can be used in a limited vocabulary context to reconstruct digitized speech with noisy, erroneous, and occluded or silenced segments.

Aircraft sometime crash in forested areas without leaving an easily detectable entry point or other indication that a crash has occurred. The wreckage can be completely obscured by overhead vegetation, which makes locating efforts, by air and ground search teams nearly impossible. In addition, there are many times, for many reasons, that emergency beacons fail to operate after a crash has occurred. For these cases, NASA has been experimenting with airborne synthetic aperture radar to provide a search tool to help focus the visual searches that are initiated after an aircraft is reported missing. This paper reviews a proposed operational scenario and the elements of a prototype airborne system that could be brought into use for finding crashed aircraft obscured from view.

In developing a beaconless search and rescue capability to quickly locate small aircraft that have crashed in remote areas, NASA's Search and Rescue Synthetic Aperture Radar (SAR²) program brings together advanced polarimetric synthetic aperture radar processing, field and laboratory tests, and state-of-the-art automated target detection algorithms. The fundamental idea underlying the search and rescue (S&R) approach is use of an airborne polarimetric radar. The downed aircraft is partly composed of metal, and consists of regular geometric shapes such as flat plates, dihedrals, trihedrals, etc., which produce a polarization signature expected to be distinct from that of surrounding terrain and foliage. Onboard polarimetric SAR image formation combined with automatic image exploitation will ultimately cue the S&R team to candidate crash sites in near real-time. We empirically examine the probability of detection (P_D) and false alarm rate (FAR) for crash site detection using polarimetry to discriminate between aircraft target signatures within natural clutter. This briefing will present the latest results from the S&R Program activities, providing an update to the last program presentation to the SPIE Meeting in 1999.

In May, 1998, the NASA Search and Rescue Mission conducted a SAR crash detection test in the swampy area south and west of Virginia Beach. A number of aircraft parts were hidden in the dense foliage. The radar used was the Navy P-3 with the ERIM XLC and UHF SAR, providing fine resolution imagery with full polarimetry and an IFSAR capability. This paper reports preliminary results of this test.

Historically, SAR to EO registration accuracy has been at the multiple pixel level compared to sub-pixel EO to EO registration accuracies. This is due to a variety of factors including the different scattering characteristics of the ground for EO and SAR, SAR speckle, and terrain induced geometric distortion. One approach to improving the SAR to EO registration accuracy is to utilize the full information from multiple SAR surveys using interferometric techniques. In this paper we will examine this problem in detail with an example using ERS SAR imagery. Estimates of the resulting accuracy based on ERS are included.

In this paper we present our approach to full resolution IFSAR processing that retains full azimuth and range resolution for a significant fraction of the pixels. This is accomplished through the use of statistical analysis and nonlinear smoothing in conjunction with limited spectral extrapolation. In our approach, a significant fraction of the pixels retain their original phase values all of the way through the processing. At the same time, we are able to reduce the number of residues by several orders of magnitude. An additional benefit of this approach is the ability to detect and discriminate between complete decorrelation due to large bodies of water and partial decorrelation due to foliage. The approach will be presented along with examples based on ERS tandem pair data.

The most important parameter in search and rescue is the time it takes to locate the downed aircraft and rescue the survivors. The resulting requirement for wide-area coverage, fine resolution, and day-night all-weather operation dictates the use of a synthetic aperture radar (SAR) sensor. The time urgency combined with the high data volume leads to the need for a new type of data compression. This paper presents and evaluates candidate compression algorithms for SAR raw phase history and for SAR complex imagery.

There are a number of methods well documented in the literature for increasing the resolution of an image product beyond the resolution limit of the EO or radar sensor that acquired the data. This paper reports on an investigation of these techniques for 'super-resolution' and their applicability to Search and Rescue Synthetic Aperture Radar (SAR²).

The potential of airborne SAR to support the search and rescue mission needs to be investigated. Interferometric SAR (IFSAR) is to process P-3 airborne SAR data to evaluate products such as Coherent Change Detection (CCD) and Digital Elevation Models (DEM). The most crucial step in this process is the precise registration of the two SAR images obtained from separate passes. This paper presents a new technique for this registration step.

It is of crucial interest for computer controlled camera systems to exhibit what is possessed by Human Vision System (HVS), which demonstrates high efficiency and effectiveness in carrying out routine perceptive tasks. To that end a wealth of different variable resolution methods have been developed to emulate the unique structure of the HVS in the research discipline of active vision. In our previous work we have developed the Foveate Wavelet Transform (FWT) as a new variable resolution representation with the benefits such as close emulation of the HVS, data reduction, and linearity preservance. Scan path is employed by the HVS to observe a scene which is larger than its field of view. An intelligent scanning scheme is developed in this paper in an effort to emulate the scan path of the HVS by taking advantage of the unique feature of the FWT. With this new scanning scheme, an ad hoc strategy is adopted to determine the next fixation point based on both a prescribed scanning order and the instant interest of the point due to its spatial and temporal properties. Preliminary experiments post encouraging performance of this proposed scanning scheme.

Image registration, i.e. correct mapping of images obtained from different sensor readings onto common reference frame, is a critical part of multi-sensor ATR/AOR systems based on readings from different types of sensors. In order to fuse two different sensor readings of the same object, the readings have to be put into a common coordinate system. This task can be formulated as optimization problem in a space of all possible affine transformations of an image. In this paper, a combination of heuristic methods is explored to register gray- scale images. The modification of Genetic Algorithm is used as the first step in global search for optimal transformation. It covers the entire search space with (randomly or heuristically) scattered probe points and helps significantly reduce the search space to a subspace of potentially most successful transformations. Due to its discrete character, however, Genetic Algorithm in general can not converge while coming close to the optimum. Its termination point can be specified either as some predefined number of generations or as achievement of a certain acceptable convergence level. To refine the search, potential optimal subspaces are searched using more delicate and efficient for local search Taboo and Simulated Annealing methods.

We describe how to teach deformable models (snakes) to find object boundaries based on user-specified criteria, and we present a method for evaluating which criteria work best. These methods prove indispensable in abdominal CT images. Further work is needed in heart ultrasound images. The methods apply in any domain with consistent image conditions characterizing object boundaries, for which automated identification is nontrivial, perhaps due to interfering detail. A traditional strongest-edge-seeking snake fails to find an object's boundary when the strongest nearby image edges are not the ones sought. But we show how to instead learn, from training data, the relation between a shape and any image feature, as the probability distribution (PDF) of a function of image and shape. An important but neglected task has always been to select image qualities to guide a model. Because success depends on the relation of objective function (PDF) output to shape correctness, it is evaluated using a sampling of ground truth, a random model of the range of shapes tried during optimization, and a measure of shape closeness. The test results are evaluated for incidence of 'false positives' (scoring better than ground truth) versus incorrectness, and for the objective function's monotonicity with respect to incorrectness. Monotonicity is measured using correlation coefficient and using the newly introduced distance from closest increasing function. Domain-dependent choices must be tested. We analyze several Gaussian models fitting image intensity and perpendicular gradient at the object boundary, as well as the traditional sum of gradient magnitudes. The latter model is found inadequate in our domains; some of the former succeed.

This paper describes a hierarchical image registration algorithm to infer the perspective transformation that best matches a pair of images. This work estimates the perspective parameters by approximating the transformation to be piecewise affine. We demonstrate the process by subdividing a reference image into tiles and applying affine registration to match them in the target image. The affine parameters are computed iteratively in a coarse-to-fine hierarchical framework using a variation of the Levenberg-Marquadt nonlinear least squares optimization method. This approach yields a robust solution that precisely registers image tiles with subpixel accuracy. The corresponding image tiles are used to estimate a global perspective transformation. We demonstrate this approach on pairs of digital images subjected to large perspective deformation.

This automatic document processing system proceeds from scanning a given paper-document into the system, automatic recognizing the document layout structure, classifying it as a particular document type, which is characterized in terms of attributes to form a frame template, and extracting the pertinent information from the document to form its corresponding frame instance, which is an effective digital form of the original document. The key attribute of the system is that it is a general-purpose system, which can be adapted easily to any application domains. A segmentation method based on the 'logical closeness' is proposed. A novel and natural representation of document layout structure -- Labeled Directed Weighted Graph (LDWG) and a methodology of transforming document segmentation into LDWG representation are described. To classify a given document, we compare its layout structure with the sample layout structures of various document types prestored in the knowledge base and then use logical structure to verify the initial matching from the first step. There is a weight associated with each component of the layout structure. During the learning stage, the system can adjust the weights automatically based on the human being's correction. Modified Perceptron Learning Algorithm (PLA) is applied.

The results of a case study about the application of an advanced method for automatic target recognition to infrared imagery taken from police helicopter missions are presented. The method consists of the following steps: preprocessing, classification, fusion, postprocessing and tracking, and combines the three paradigms image pyramids, neural networks and bayesian nets. The technology has been developed using a variety of different scenes typical for military aircraft missions. Infrared cameras have been in use for several years at the Bavarian police helicopter forces and are highly valuable for night missions. Several object classes like 'persons' or 'vehicles' are tested and the possible discrimination between persons and animals is shown. The analysis of complex scenes with hidden objects and clutter shows the potentials and limitations of automatic target recognition for real-world tasks. Several display concepts illustrate the achievable improvement of the situation awareness. The similarities and differences between various mission types concerning object variability, time constraints, consequences of false alarms, etc. are discussed. Typical police actions like searching for missing persons or runaway criminals illustrate the advantages of automatic target recognition. The results demonstrate the possible operational benefits for the helicopter crew. Future work will include performance evaluation issues and a system integration concept for the target platform.

This paper describes a conceptual real-time systems approach to LADAR automatic target recognition (ATR). Previous work has demonstrated the viability of utilizing correlation filters derived from synthetic models for detection and recognition of mobile targets in Laser Radar (LADAR) sensor images. The distance correlation classifier filter (DCCF) provides a unique potential for reducing throughput while preserving performance. The application of this concept to a real-time system, however, involves refinement, trade studies, and optimization. Refinements in the correlation filter are discussed and evaluated in terms of performance, throughput, and memory. Preliminary performance results for several mobile targets are presented.

Target detection and recognition are two important modules in a typical automatic target recognition (ATR) system. Usually, an automatic target detector produces many false alarms that could incur very poor recognition performance in the subsequent target recognizer. Therefore, we need a good clutter rejector to remove as many clutters as possible from the outputs of the detector, before feeding the most likely target detections to the recognizer. We investigate the benefits of using dualband forward-looking infrared (FLIR) images to improve the performance of a eigen-neural based clutter rejector. With individual or combined bands as input, we use either principal component analysis (PCA) or the eigenspace separation transform (EST) to perform feature extraction and dimensionality reduction. The transformed data is then fed to an MLP that predicts the identity of the input, which is either a target or clutter. We devise an MLP training algorithm that seeks to maximize the class separation at a given false-alarm rate, which does not necessarily minimize the average deviation of the MLP outputs from their target values. Experimental results are presented on a dataset of real dualband images.

Model-based automatic target recognition (ATR) using forward- looking infrared (FLIR) imagery, and using FLIR imagery combined with cues from a synthetic aperture radar (SAR) system, has been successfully demonstrated in the laboratory. For the laboratory demonstration, FLIR images, platform location, sensor data, and SAR cues were read in from files stored on computer disk. This ATR system, however, was intended to ultimately be flown in a fighter aircraft. We discuss the transition from laboratory demonstration to flight demonstration for this system. The obvious changes required were in the interfaces: the flight system must get live FLIR imagery from a sensor; it must get platform location, sensor data, and controls from the avionics computer in the aircraft via 1553 bus; and it must get SAR cues from the on-board SAR system, also via 1553 bus. Other changes included the transition to rugged hardware that would withstand the fighter aircraft environment, and the need for the system to be compact and self-contained. Unexpected as well as expected challenges were encountered. We discuss some of these challenges, how they were met, and the performance of the flight-demonstration system.

In this article, we present two mathematical paradigms for clutter modeling. Both paradigms pose clutter modeling as a statistical inference problem, and pursue probabilistic models for characterizing observed training images. The two paradigms differ in the forms (or families) of models that they choose and in their philosophical assumptions on real world clutter patterns. The first paradigm studies descriptive models, such as Markov random field (MRF) models and the minimax entropy models (Zhu, Wu, and Mumford 1997). In this modeling paradigm, image features are first extracted from images, and statistics of these features are calculated. The latter define an image ensemble-called the Julesz ensemble which is an equivalence class where all images share the same feature statistics. For any large images from this ensemble, a local patch given its boundary condition is then Gibbs (or MRF) models. We shall review the recent conclusions about ensemble equivalence studied in (Wu, Zhu and Liu, 1999). The second paradigm studies generative model, such as the random collage model (Lee and Mumford, 1999). In contrast to a descriptive model, a generative model introduces hidden variables which are assumed to be the underlying causes producing the observed image. For example, trees and rock for clutter. The learning process makes inference about the hidden variables. We shall discuss a texton model for clutter and effective Markov chain Monte Carlo methods for stochastic inference. We shall also reveal the deep relationship between the two modeling paradigm.

This article presents a mathematical paradigm called Data Driven Markov Chain Monte Carlo (DDMCMC) for effective stochastic inference in the Bayesian framework. We apply the DDMCMC paradigm to two typical problems in image analysis: object recognition and image segmentation. In both problems, the solution spaces are not only high dimensional but heterogeneously-structured in the sense that they are composed of many subspaces of varying dimensions. Each of the subspace is product of what we called the object spaces. The latter is further decomposed into the so-called atomic spaces. The DDMCMC paradigm simulates Markov chains for exploring the solution spaces using both jump and diffusion dynamics. Unlike traditional MCMC algorithms, the DDMCMC paradigm utilizes data driven (or bottom-up) techniques, such as Hough transform, edge detection, and color clustering, to design effective transition probabilities for Markov chain dynamics. This drastically improves the effectiveness of traditional MCMC algorithms in terms of two standard metrics: 'burn-in' period and 'mixing' rate. The article proceeds in three steps. Firstly, we analyze the structures of the solution space (Omega) for the two tasks. Secondly, we study how data-driven techniques are utilized to compute importance proposal probabilities in the solution spaces, the object spaces and atomic spaces. These proposal probabilities are expressed in non-parametric form using weighted samples or particles. Thirdly, we design Markov chains to travel in such heterogeneous structured solution space as an ergodic and reversible process. The paper first review the DDMCMC theory using a simple object recognition problem -- the (psi) -- world reported in [14], then we briefly introduce the results on image segmentation.

Many problems in image analysis and ATR can be formulated as Bayesian inference. In this paper we present a unified framework to quantify performance (i.e. the accuracy and uncertainty of inference) in terms of Bayesian decision theory. We demonstrate that existing work on image analysis and ATR performance can be summarized in terms of these concepts. This includes performance measures such as signal- to-noise ratios, Cramer-Rao lower bounds, Hilbert-Schmidt bounds, and ROC curves. Secondly, we describe how recent work by the authors on order parameters can be reformulated within this framework. This includes analyzing how phase transitions can occur for target detection problems so that at critical values of the order parameters it becomes impossible to detect the target. We also analyze the case where the inference process uses weaker prior knowledge to detect the target and quantify in what situations this strategy is effective.

This paper derives bounds on the performance of statistical object recognition systems, wherein an image of a target is observed by a remote sensor. Detection and recognition problems are modeled as composite hypothesis testing problems involving nuisance parameters. We develop information- theoretic performance bounds on target recognition based on statistical models for sensors and data, and examine conditions under which these bounds are tight. In particular, we examine the validity of asymptotic approximations to probability of error in such imaging problems. Applications to target recognition based on compressed sensor image data are given. This study provides a systematic and computationally attractive framework for analytically characterizing target recognition performance under complicated, non-Gaussian models, and optimizing system parameters.

A method for robustly selecting reduced dimension statistics for pattern recognition systems is described. A stochastic model for each target or object is assumed parameterized by a finite dimensional vector. Data and parameter vectors are assumed to be long. As the size of these vectors increases, the performance improves to a point and then degrades; this trend is called the peaking phenomenon. A new, more robust method for selecting reduced dimension approximations is presented. This method selects variables if a measure of the amount of information provided exceeds a given level. This method is applied to distributions in the exponential family, performance is compared to other methods, and an analytical expression for performance is asymptotically approximated. In all cases studied, performance is better than with other known methods.

The field of moving target Automated Target Recognition (ATR) relies on the exploitation of one-dimensional high range resolution (HRR) profiles. Individual profiles can contain a large amount of target information; however, the evidence from one profile is generally not sufficient to reliably classify a Ground Moving Target Indicator (GMTI) target. When multiple looks are correctly combined, classification accuracy can improve dramatically. At X-band, HRR profiles of typical ground vehicles decorrelate for aspect angle changes greater than 0.1 degree, thus, all looks in a practical system are independent. From the ATR perspective, the challenge is one of correctly associating HRR profiles from one look to the next. If the problem is examined from the opposite point of view, the ATR evidence can greatly improve the association accuracy of a tracker above and beyond that of kinematics. This ATR information assists tracking in regimes of high traffic density or low revisit rates through better association of the high-value targets from one epoch to the next. In this paper, we present a new HRR-aided tracker. The performance of this tracker will be characterized in a simulation and compared to the performance of a purely kinematic tracker. These results show that HRR-aided tracking can tolerate at least an order of magnitude higher traffic density than trackers functioning on kinematics alone. This improvement in performance is reduced, but not eliminated, if the additional radar resources for HRR are considered.

This paper presents an algorithm to estimate motion vectors from Polarimetric IR data sequences. In the proposed algorithm, based on the I, P, and (psi) frames of a PIR sequence, motion estimation is formulated as a problem of obtaining the Maximum A Posteriori in the Markov Random Field (MAP-MRF). An optimization method based on the Mean Field Theory (MFT) is chosen to carry out the MAP search. The estimation of motion vectors is modeled by two MRF's, namely, motion vector field and unpredictable field. A truncation function is employed to handle the discontinuity between motion vectors on neighboring sites. In this algorithm, a 'double threshold' step is first applied to partition the sites into three regions, whereby the ensuing MFT-based step for each MRF is performed on one or two of the three regions. With this algorithm, no significant difference exists between the block-based and pixel-based MAP searches any more. Consequently, a good compromise between precision and efficiency can be obtained with ease.

A common image processing problem is determining the location of an object using a template when the size and rotation of the object are unknown. In the case of known geometric parameters, it is possible to use an impulse reconstruction technique to determine object location. In the case of unknown parameters, we show that localization is possible by computing a likelihood surface for a dense sampling of the size and rotation space. However, the surface produced is not amenable to conventional minimization methods due to local minima and regions of small or zero gradient. Using a smooth approximate template, we can overcome these difficulties at the expense of estimation accuracy. We therefore demonstrate a technique which employs a library of templates starting from the smooth approximation and adding detail until the exact template is reached. Successively estimating the geometric parameters using these templates achieves the accuracy of the exact template while remaining within a well-behaved 'bowl' in the search space which allows standard minimization techniques to be used.

SAR resolution and polarization performance studies for ATR algorithms have been the source of recent attention. Thorough investigations are often hindered by the lack of rigorously consistent high-resolution full-polarimetric signature data for a sufficient number of targets across requisite viewing angles, articulations and environmental conditions. While some evaluative performance studies of high-value structures and conceptual radar systems may be effectively studied with limited field radar data, to minimize signature acquisition costs, pose-independent studies of ATR algorithm are best served by signature libraries fashioned to encompass the complexity of the collection scenario. In response to the above requirements, the U.S. Army's National Ground Intelligence Center and Targets Management Office originated, sponsored, and directed a signature project plan to acquire multiple target signature data at Eglin, AFB using a high resolution full-polarimetric Ka-band radar. TMO and NGIC have sponsored researchers at both the Submillimeter-Wave Technology Laboratory and Simulation Technologies to analyze the trade-off between signature resolution and polarimetric features (ongoing research) of this turntable data. The signature data was acquired at five elevations spanning 5 degree to 60 degree for a T-72M1, T-72B, M1, M60-A3 and one classified vehicle. Using signal processing software established in an NGIC/STL-based signature study, researchers executed an HRR and ISAR cross-correlation study involving multiple resolutions to evaluate peak performance levels and to effectively understand signature requirements through the variability of multiple target RCS characteristics. The signature-to-signature variability quantified on the four unclassified MBTs is presented in this report, along with a description and examples of the signature analysis techniques exploited. This signature data is available from NGIC/TMO on request for Government Agencies and Government Contractors with an established need-to-know.

A goal of super-resolution, in addition to improving probability of correct classification (Pcc) in automatic target recognition systems, is to reduce radar resource requirements in achieving a given Pcc. These studies address the MIT Lincoln Laboratory 1-D template-based ATR algorithm that was developed and tested on super-resolved high range resolution (HRR) profiles formed from synthetic aperture radar (SAR) images of targets taken from the Moving and Stationary Target Acquisition and Recognition (MSTAR) data set. Previous studies on HRR ATR demonstrated encouraging results for recognition of stationary targets from their HRR profiles, although the low probability of correct classification dictates a large margin of improvement in Pcc is needed before the system can be operational. In this work, a super- resolution technique known as High Definition Vector Imaging (HDVI) is applied to the HRR profiles before the profiles are passed through the ATR classification. The new 1-D ATR system using super-resolved HRR demonstrates significantly improved target recognition compared to previous 1-D ATR systems that use conventional image processing techniques. This paper discusses the improvement in HRR ATR performance in terms of radar resource requirements as a result of applying HDVI.

There has been much attention given to the characterization of canonical scattering phenomena in complex synthetic aperture radar (SAR) imagery for the purpose of classification. These features are often extracted using greedy algorithms, such as Matching Pursuits, which attempt to extract a low-dimensional representation of a given SAR image using a prescribed feature dictionary. As results to date have been predominantly anecdotal in nature, it is of interest to assess the utility of these techniques in full automated target classification applications. In this investigation, we will focus on the potential incorporation of such techniques into automatic target recognition (ATR) systems. The primary issues addressed in this paper include robust feature extraction techniques and the development of effective likelihood-of-match metrics operating in feature space. A specific implementation is presented where robust feature extraction is achieved via a new technique called Directed Pursuits. Directed Pursuits constitutes a feature extraction process from a test image driven by a feature decomposition previously obtained via Matching Pursuits from a training image. Directed Pursuits additionally allows for spatial flexibility in the test extraction process, representative of natural intra-class variations exhibited by real target classes. Likelihood metrics will be motivated and described, and associated results presented and interpreted in the context of both airborne and compact range X-band SAR data.

The M-ary sequential probability ratio test MSPRT is used to recognize unknown non-cooperative radar targets. Radar returns representing the unknown target backscatter coefficients are tested sequentially using MSPRT. At each stage of the recognition process all observations are used in MSPRT, if no identification decision can be made, additional information is requested and MSPRT is implemented again. The goal is either to minimize the number of observations needed to identify an unknown target assuming a certain predetermined error probability, or to minimize the probability of error assuming a predetermined maximum number of observations. The experimental phase of this study involves radar cross-section signatures of four commercial aircraft models recorded in a compact range environment. Scenarios representing various degrees of azimuth uncertainty are examined in this paper. In all cases, it is assumed that the unknown target is corrupted with additive white Gaussian noise.

Automatic Target Recognition (ATR) of moving targets has recently received increased interests. High Range Resolution (HRR) radar mode provides a promising approach which relies on processing high-resolution 'range profiles' over multiple look angles. To achieve a robust, reliable and cost effective approach for HRR-ATR, a model-based approach is investigated in this paper. A subset of the Moving and Stationary Target Acquisition and Recognition (MSTAR) data set was used to study robustness and sensitivity issues related to 1D model-based ATR development and performance. The model is built based on the statistic analysis of the training data and the dependence of the HRR signature on the azimuth is considered. The dependence is approximated by a linear regression algorithm to construct the templates of targets, which gives this approach the name of piecewise linear approach (PWL). Compared with the 1D model-based ATR approach developed by the Wright Laboratory, results are presented demonstrating an increase of about 10% in the correct identification probability of known targets when declaration probability P_dec is above 85% while maintaining a low time-cost.

Automated target recognition (ATR) is a problem of great importance in a wide variety of applications: from military target recognition to recognizing flow-patterns in fluid- dynamics to anatomical shape-studies. The basic goal is to utilize observations (images, signals) from remote sensors (such as videos, radars, MRI or PET) to identify the objects being observed. In a statistical framework, probability distributions on parameters representing the object unknowns are derived an analyzed to compute inferences (please refer to [1] for a detailed introduction). An important challenge in ATR is to determine efficient mathematical models for the tremendous variability of object appearance which lend themselves to reasonable inferences. This variation may be due to differences in object shapes, sensor-mechanisms or scene- backgrounds. To build models for object variabilities, we employ deformable templates. In brief, the object occurrences are described through their typical representatives (called templates) and transformations/deformations which particularize the templates to the observed objects. Within this pattern-theoretic framework, ATR becomes a problem of selecting appropriate templates and estimating deformations. For an object (alpha) (epsilon) A, let I^(alpha ) denote a template (for example triangulated CAD-surface) and let s (epsilon) S be a particular transformation, then denote the transformed template by sI^(alpha ). Figure 1 shows instances of the template for a T62 tank at several different orientations. For the purpose of object classification, the unknown transformation s is considered a nuisance parameter, leading to a classical formulation of Bayesian hypothesis- testing in presence of unknown, random nuisance parameters. S may not be a vector-space, but it often has a group structure. For rigid objects, the variation in translation and rotation can be modeled through the action of special Euclidean group SE(n). For flexible objects, such as anatomical shapes, higher-dimensional groups such as a diffeomorphisms are utilized.

In this paper, we discuss stability and accuracy of target detection process in infrared images, using continuous features of the target or target candidates, and we improve the accuracy of target detection by optimizing an evaluation function. This process carries out parallel image process. The one, Two-dimensional Constant False Alarm Rate (CFAR) process reduces background clatter. The other, Motion Vector Process detects moving target. In addition, Combined Target Detection Process improves the accuracy of target detection using features from two different processes. Continuous and stable data measurement is necessary to improve the accuracy of target detection. But measurement data have noises and fluctuations by change of the environment. In this case, we use continuous features to stable detection. On the other hand, optimizing weight vectors in evaluation function is necessary to improve target detection. But we have to deal with large number of parameters. In this optimization named Combined Target Detection Process, we use Genetic Algorithms (GA) to get a global optimum of parameters. This process is useful for outdoor surveillance systems, intelligent transport systems (ITS) and so on.