This work deals with the development of Kronecker products algorithms for fast synthetic aperture radar (SAR) image formation operations. A methodology has been developed to serve as a tool aid in the analysis, design, and efficient implementation of one-dimensional and two-dimensional fast Fourier transform (FFT) algorithms prevalent in SAR image formation operations with the idea in mind of reducing the computational effort and improving the hardware implementation process for real time on board SAR imaging applications. Kronecker products algebra, a branch of finite dimensional multilinear algebra, has been demonstrated to be a useful tool aid in the development of fast algorithms for unitary transformations and in the identification of similarities and differences among FFT computational frameworks.

Statistical orthogonality provides a mathematical basis for imaging scattering data with an inversion algorithm that is both robust and economic. The statistical technique is based on the approximate orthogonality of vectors whose elements are exponential functions with imaginary arguments and random phase angles. This orthogonality allows one to image radar data without first inverting a matrix whose dimensionality equals or exceeds the number of pixels or voxels in the algorithmic image. Additionally, statistical-based methods are applicable to data sets collected under a wide range of operational conditions, e.g., the random flight paths of the curvilinear SAR, the frequency-hopping emissions of ultra- wideband radar, or the narrowband data collected with a bistatic radar. The statistical approach also avoids the often-challenging and computationally intensive task of converting the collected measurements to a data format that is appropriate for imaging with a fast Fourier transform (FFT) or fast tomography algorithm (FTA), e.g., interpolating from polar to rectangular coordinates, or conversely.

Radar imaging is often posed as a problem of estimating deterministic reflectances observed through a linear mapping and additive Gaussian receiver noise. We consider an alternative view which considers the reflectances themselves to be a realization of a random process; imaging then involves estimating the parameters of that underlying process. Purely diffuse radar targets are modeled by a zero-mean Gaussian process, while specular targets introduce an additive component with fixed amplitude and random uniform phase. When conventional stepped frequency waveforms are employed, the linear mapping amounts to a Fourier transform, and parameter estimation is straightforward. If more complicated waveforms are employed, maximum-likelihood parameter estimates cannot be readily computed analytically; hence, we explore an iterative expectation-maximization algorithm proposed by Snyder, O'Sullivan, and Miller. Although this algorithm was designed for diffuse radar imaging, arguments based on the central limit theorem and computational experiments support its applicability to the specular case. The resulting estimates tend to be unacceptably rough due to the ill-posed nature of maximum-likelihood estimation of functions from limited data, so some kind of regularization is needed. We explore penalized likelihoods based on entropy functionals, a roughness penalty proposed by Silverman, and an information-theoretic formulation of Good's roughness penalty.

Due to the ordinary low resolution radar can not distinguish the radar target in both range and azimuth. If we apply the technology of inverse synthetic aperture radar (ISAR) to resolve the difference among Doppler frequency of the scatters on the target, we can obtain a fine resolution cross-range image. The cross-range scale depends on both radar wavelength and rotating angle of target relative to radar-line-of-sight (RLOS) during the coherent accumulation. The former is known while the latter is difficult to determine especially in the case of ISAR. But we must investigate the method of cross- range scaling of low-resolution radar, as it is very important to radar target classification and recognition. In this paper, a new approach is proposed which is based on the principle of interferometric inverse synthetic aperture. We can calculate the phase difference of some scatters between two instant cross-range images by two antenna which are placed on one level, adding the range between the two radar and the range of the target, and then absolute cross ranges of some dominant scatters are obtained. We apply the proposed algorithm to the emulational data of two antennae. The processing results show that the proposed method is correct and effective.

One of the biggest challenges for automatic target recognition (ATR) methods is the accurate and efficient extraction of features from synthetic aperture radar (SAR) images. We have recently developed a new SAR image formation technique which is recognition-oriented in the sense that it enhances features in the scene which we believe are important for recognition purposes. In this paper, we evaluate the performance of the images produced by this technique in terms of preserving and enhancing these features. The findings of our analysis indicate that the new SAR image formation method provides images with higher resolution of scatterers, and better separability of different regions as compared to conventional SAR images.

A nonlinear reconstruction scheme based on the Distorted Born Iterative Method (DBIM) applied to metallic scatterers is presented to solve two-dimensional inverse scattering problems. By making use of half-quadratic regularization the inherently ill-posed problem of nonlinear inverse scattering can be alleviated and edges of the scatterer's profile function simultaneously be preserved. Reconstruction results are found by the minimization of a cost function. A bistatic simulated experimental set-up is proposed in angular and frequency diversity. A Multilevel Fast Multipole Algorithm (MLFMA) in combination with a conjugate gradient (CG) scheme is introduced to solve the forward as well as the inverse scattering problem involved in the DBIM algorithm. Through the use of MLFMA for the forward and for the inverse solver, the computational complexity per CG iteration is of order O(N log N) compared to O(N²) in a standard implementation without the MLFMA. Numerical reconstruction results obtained from synthetic scattering data are presented.

This paper address the problem of processing an X-band SAR database that was originally intended for processing via a polar format imaging algorithm. In our approach, we use the approximation-free SAR wavefront reconstruction. For this, the measured and motion compensated phase history (polar format) data are processed in a multi-dimensional digital signal processing algorithm that yields alias-free slow-time samples. The resultant database is used for wavefront image formation. The X-band SAR system also provides a two channel along-track monopulse database. The alias-free monopulse SAR data are used in a coherent signal subspace algorithm for Ground Moving Target Indication (GMTI). Results are provided.

Interferometric SAR (IFSAR) can be shown to be a special case of 3-D SAR image formation. In fact, traditional IFSAR processing results in the equivalent of merely a super- resolved, under-sampled, 3-D SAR image. However, when approached as a 3-D SAR problem, a number of IFSAR properties and anomalies are easily explained. For example, IFSAR decorrelation with height is merely ordinary migration in 3-D SAR. Consequently, treating IFSAR as a 3-D SAR problem allows insight and development of proper motion compensation techniques and image formation operations to facilitate optimal height estimation. Furthermore, multiple antenna phase centers and baselines are easily incorporated into this formulation, providing essentially a sparse array in the elevation dimension. This paper shows the Polar Format image formation algorithm extended to 3 dimensions, and then proceeds to apply it to the IFSAR collection geometry. This suggests a more optimal reordering of the traditional IFSAR processing steps.

The paper describes a method for phase unwrapping of interferometric SAR (synthesize aperture radar). The method use edge detection techniques to find the location of the phase jump (fringe lines) in the interferogram. Here, we combine the fringe detection and tracking to improve the accuracy of the fringe location. Then, a method of the phase correcting is also given to correct phase leap. The method given the paper has advantages of stability and less error spread. Finally, by using the L band data of SIR-C of the Three-Gorge area of China, the data of phase unwrapping are obtained by the method.

Useful products generated from interferometric synthetic aperture radar (IFSAR) complex data include height measurement, coherent change detection, and classification. The IFSAR coherence is a spatial measure of complex correlation between two collects, a product of IFSAR signal processing. A tacit assumption in such IFSAR signal processing is that the terrain height is constant across an averaging box used in the process of correlating the two images. This paper presents simulations of IFSAR coherence if two target with different heights exist in a given correlation cell, a condition in IFSAR collections produced by layover. It also includes airborne IFSAR data confirming the simulation results. The paper concludes by exploring the implications of the results on IFSAR height measurements and classification.

In practice, a synthetic aperture radar (SAR) reconstructs the complex reflectivity function of a scene, modulated by phase terms that capture 3-D imaging geometry. INSAR (interferometric SAR) attempts to obtain the geometric information by interfering two images (from two antennas) to cancel the same scene reflectivity and recover the scene topography transduced by the image-phase data. This approach, however, leads to a phase-unwrapping problem, which causes ambiguities in estimates of elevation. The phase-unwrapping problem can be solved in a pointwise fashion by using more than two antennas. This approach can effectively prevent error propagation which occurs in traditional phase-unwrapping algorithms. In this work, we study the optimal antenna spacings for pointwise terrain height estimation. In particular, we start from the maximum likelihood estimates of the phase using neighborhood pixels collected by any pair of antennas. The phase estimation noise is approximated as Gaussian with variance prescribed by the Cramer-Rao lower bound on the phase estimate. The ambiguous terrain height derived from any pair of antennas is modeled by a periodic waveform with each period having an approximately Gaussian shape. For multiple pairs of antennas, the corresponding functions describing the ambiguous elevation have different periods, which acts to help resolve the ambiguity. We derive and analyze the ML estimate of elevation at each scene point using multiple pairs of antennas. For the three-antenna case, by analyzing the tradeoff between cycle errors and measurement errors, a closed-form formula approximating the mean squared error (MSE) of the estimated terrain height is derived as a function of antenna spacing. By minimizing the MSE, we determine the optimal antenna spacing. The algorithm is tested with simulated data.

We study the segmentation of SAR imagery using wavelet-domain Hidden Markov Tree (HMT) models. The HMT model is a tree- structured probabilistic graph that captures the statistical properties of the wavelet transforms of images. This technique has been successfully applied to the segmentation of natural texture images, documents, etc. However, SAR image segmentation poses a difficult challenge owing to the high levels of speckle noise present at fine scales. We solve this problem using a 'truncated' wavelet HMT model specially adapted to SAR images. This variation is built using only the coarse scale wavelet coefficients. When applied to SAR images, this technique provides a reliable initial segmentation. We then refine the classification using a multiscale fusion technique, which combines the classification information across scales from the initial segmentation to correct for misclassifications. We provide a fast algorithm, and demonstrate its performance on MSTAR clutter data.

We present a new approach to SAR image segmentation based on a Poisson approximation to the SAR amplitude image. It has been established that SAR amplitude images are well approximated using Rayleigh distributions. We show that, with suitable modifications, we can model piecewise homogeneous regions (such as tanks, roads, scrub, etc.) within the SAR amplitude image using a Poisson model that bears a known relation to the underlying Rayleigh distribution. We use the Poisson model to generate an efficient tree-based segmentation algorithm guided by the minimum description length (MDL) criteria. We present a simple fixed tree approach, and a more flexible adaptive recursive partitioning scheme. The segmentation is unsupervised, requiring no prior training, and very simple, efficient, and effective for identifying possible regions of interest (targets). We present simulation results on MSTAR clutter data to demonstrate the performance obtained with this parsing technique.

This paper presents a technique for detection of roads in a spaceborne SAR image using a genetic algorithm. Roads in a spaceborne SAR image can be modelled as curvilinear structures with some thickness. Curve segments, which represent candidate positions of roads, are extracted from the image using a curvilinear structure detector, and roads are detected accurately by grouping those curve segments. For this purpose, we designed a grouping method based on a genetic algorithm (GA), which is one of the global optimization methods, combined perceptual grouping factors with it, and tried to reduce its overall computational cost by introducing an operation of thresholding and a concept of region growing. To detect roads more accurately, postprocessing, including noisy curve segment removal, is performed after grouping. We applied our method to ERS-1 SAR images that have a resolution of about 30 meters, and the experimental results show that our method can detect roads accurately, and is much faster than a globally applied GA approach.

The problem of estimating the dimension parameters of rectangular and circular cavities from scattering data is addressed in this paper. By incorporating the modal solutions of the waveguides in the Complex Phase History Signal model, the Maximum-Likelihood Estimation (MLE) criterion has been developed for the rectangular cavity. It is shown that the MLE is a multi-dimensional non-linear optimization problem and optimization techniques have been proposed to address this problem. The cavity detection problem has also been formulated within the multiple Hypotheses testing framework. The proposed ML estimation approaches have been verified with simulation studies.

The complex phase history of the synthetic aperture radar (SAR) image can be broken up into a series of overlapping crossrange sub-apertures from which time-dependent information can be extracted from the scatterers that make up the image. The change in the image characteristics from one sub-aperture window to the next can be correlated to generate a stability map of the scene; i.e., areas that change rapidly as the sensor viewpoint changes can be identified. These changes in pixel characteristics over the aperture may be due to many sources: scatterers interfering with each other differently as acquisition geometry is swept out, scatterer sources being created and destroyed as the geometry changes, and aspect- dependent scatterers (e.g. dihedrals) being interrogated differently over the width of the aperture. The stability analysis reported earlier for the peak features has been extended to arbitrary pixels in the SAR image, and the application of this analysis to other features of interest can be made. An example using measured MSTAR data of the stability of a characterized scatterer from the Slicey phantom, imaged in a self-obscuring state, is presented in the context of the stability analysis. Further generalizations of the approach to polarimetric SAR are also presented.

We present a semi-parametric spectral estimation algorithm for fully polarimetric synthetic aperture radar (SAR) target feature extraction and image formation. The algorithm is based on a flexible data model that models each target scatterer as a two-dimensional complex sinusoid with arbitrary amplitude and constant phase in cross-range and with constant amplitude and phase in range. The algorithm is a relaxation-based optimization approach that minimizes a nonlinear least squares (NLS) cost function. Due to using the fully polarimetric radar measurements (HH, HV, and VV) simultaneously, the algorithm provides not only more accurate target features, but also more useful information about the target of interest than the single polarization based algorithm. The algorithm has the ability to discriminate corner reflector types by also exploiting the differences in the polarimetric scattering properties of the scatterers of the target of interest. Numerical examples are presented to demonstrate the performance of the proposed algorithm.

Scattering from man-made objects in SAR imagery often exhibit aspect and frequency dependences which are not well modeled by standard SAR imaging techniques. If ignored, these deviations may reduce recognition performance due to model mismatch, but when appropriately accounted for, these deviations can be exploited as attributes to better distinguish scatterers and their respective targets. Chiang and Moses developed an ATR system that allows the study of performance under various scatterer attributions. Kim et. al. examined a nonparametric approach for exploiting non-ideal scattering using a multi- resolution sub-aperture representation. Both of these works are extended here to examine the effect of anisotropic scattering attribution for model-based ATR. In particular, predicted and extracted peak scatterers are attributed with a discrete anisotropy feature. This feature can be obtained in a computationally efficient manner by performing a set of generalized log-likelihood ratio (GLLR) tests over a pyramidal sub-aperture representation. Furthermore, an approximate probabilistic characterization of the feature set allows for a natural inclusion into the approach of Chiang and Moses which will be used to evaluate the benefit of our attribution to the X-band MSTAR data and infer the phenomenology behind anisotropic scattering.

The inherent poor resolution in several practical SAR image acquisition operations usually demands the processing of acquired imagery data for resolution enhancement before the data can be used for any target surveillance and classification purposes. Among the available restoration methods, super-resolution algorithms are primarily directed to re-creating some of the image spectral content (specifically the high frequency components that are responsible for the higher resolution) that is lost due to diffraction-limited sensing operations. While several different approaches for image super-resolution have been proposed in the recent past, iterative processing algorithms developed using statistical optimization methods have attained considerable prominence. In particular, a set of algorithms that attempt to maximize the likelihood of the image estimate by employing distribution functions that are simple to handle within an optimization framework have been shown to yield remarkable spectral extrapolation, and hence super-resolution, performance. Notwithstanding the powerful capabilities of these algorithms, a direct processing of SAR images with these could lead to generation of processing artifacts. Since the fundamental idea underlying the restoration and super-resolution processing is an intelligent utilization of a priori information (about the object or scene imaged), improved performance can be realized by incorporating constraint set modeling approaches in the maximization of the likelihood function. This paper will discuss the modeling of specific constraint sets and the design of processing algorithms using convex set models of the a priori information. The restoration and super-resolution performance of these algorithms will be described. We also discuss a few hybrid algorithms that combine the strong points of the Maximum Likelihood (ML) algorithm and the convex set based processing methods. Quantitative performance evaluation of these algorithms in processing SAR images is also given by application of these methods to MSTAR data.

A robust autofocus approach, referred to as AUTOCLEAN (AUTOfocus via CLEAN), is proposed for the motion compensation in ISAR (inverse synthetic aperture radar) imaging of moving targets. It is a parametric algorithm based on a very flexible data model which takes into account arbitrary range migration and arbitrary phase errors across the synthetic aperture that may be induced by unwanted radial motion of the target as well as propagation or system instability. AUTOCLEAN can be classified as a multiple scatterer algorithm (MSA), but it differs considerably from other existing MSAs in several aspects: (1) dominant scatterers are selected automatically in the two-dimensional (2-D) image domain; (2) scatterers may not be well-isolated or very dominant; (3) phase and RCS (radar cross section) information from each selected scatterer are combined in an optimal way; (4) the troublesome phase unwrapping step is avoided. AUTOCLEAN is computationally efficient and involves only a sequence of FFTs (fast Fourier Transforms). Another good feature associated with AUTOCLEAN is that its performance can be progressively improved by assuming a larger number of dominant scatterers for the target. Hence it can be easily configured for real-time applications including, for example, ATR (automatic target recognition) of non-cooperative moving targets, and for some other applications where the image quality is of the major concern but not the computational time including, for example, for the development and maintenance of low observable aircrafts. Numerical and experimental results have shown that AUTOCLEAN is a very robust autofocus tool for ISAR imaging.

In this paper, the phase of a radar range profile is shown to contain valuable information for inverse scattering problems. A physics-based high-frequency parametric model is adopted for the radar backscatter, and information is quantified using the variance of parameters estimated from noisy radar range profiles. Through analysis of the Fisher information matrix, phase is observed to yield up to a factor of ten increase in achievable resolution; moreover, phase is shown to allow reliable discrimination of frequency-dependent scattering behaviors. Results are confirmed using measured radar imagery from a 2-inch resolution X-band system.

The Army Research Laboratory (ARL) has been developing ultra wideband (UWB), ultra wide angle radar technology to meet warfighter requirements to detect concealed targets (such as tactical vehicles under foliage). Experiments undertaken by ARL and others using testbed radar's (such as ARL's BoomSAR) have shown significant potential for detecting hidden targets. Initial evaluations have concentrated on identifying the 'contrast' ratios for desired targets versus average background. In more recent work, we have begun to evaluate specific angle, frequency, and/or polarization-based scattering properties of targets and clutter to isolate discrimination features for use in automatic target detection and cuing (ATD/C) algorithms (see reference 1). Though promising, much of this work has been ad hoc and based on small data sets that have only recently become available. To complement the measurements and analysis effort under way at ARL, our team is also developing high-fidelity electromagnetic models of targets and certain classes of clutter to gain a physics-based insight into robust discrimination techniques. We discuss recent analysis of both EM model results as well as a unique inverse synthetic aperture radar (ISAR) collection undertaken at Aberdeen Proving Ground (APG). By creating a phenomenological framework for explaining and/or describing target and/or clutter backscatter behavior and comparing it with measured field data, we can develop detection strategies inspired by the unique physics of low-frequency radar. Finally, we suggest one such detection paradigm.

In the frame of Bayes statistic criterion of optimization the processing algorithms and the structures of radiometric systems for wide- and superwide band electro-magnetic fields spatial-temporal processing are developed. The methods of imaging in multi-beam systems and in systems, that form coherence functions of decorrelated processes. The algorithms has been developed using proposed by authors VF-transforms and proved by them the theorem, that are generalization of Fourier transforms and Van Zittert-Zernike theorem, respectively, for the case of wide- and superwide band wave fields spectral analysis.

This paper presents a novel scheme to detect and discriminate landmines from other clutter objects during the image formation process for ultra-wideband (UWB) synthetic aperture radar (SAR) systems. By identifying likely regions containing the targets of interest, i.e., landmines, it is possible to speed up the overall formation time by pruning the processing to resolve regions that do not contain targets. The image formation algorithm is a multiscale approximation to standard backprojection known as the quadtree that uses a 'divide-and- conquer' strategy. The intermediate quadtree data admits multiresolution representations of the scene, and we develop a contrast statistic to discriminate structured/diffuse regions and an aperture diversity statistic to discriminate between regions containing mines and desert scrub. The potential advantages of this technique are illustrated using data collected at Yuma, AZ by the ARL BoomSAR system.

Radar scattering from an illuminated object is often dependent on target-sensor orientation. In synthetic aperture radar (SAR) imagery, the aspect dependence of the target over the aperture is lost during image formation. To recover this directional dependence, we post-processes the SAR imagery to generate a sequence of images over a corresponding sequence of subapertures. Features are extracted from the sequence of subaperture images using a two-dimensional matching pursuits algorithm. The feature statistic associated with geometrically distinct target-sensor orientations are then used to design a hidden Markov model (HMM) for the target class. This approach explicitly incorporates the sensor motion into the model and accounts for the fact that the orientation of the target is assumed to be unknown. Performance is quantified by considering the detection of tactical targets concealed in foliage.

This paper is concerned with the implementation of the SAR wavefront reconstruction algorithm on a high performance computer. For this purpose, the imaging algorithm is reformulated as a coherent processing (spectral combination) of images that are formed from a set of subapertures of the available synthetic aperture. This is achieved in conjunction with extracting the signature of a specific target region (digital spotlighting). Issues that are associated with implementing the algorithm on SMP-HPCs and DMP-HPCs are discussed. The results using the FOPEN P-3 SAR data are provided.

Many radar automatic target detection (ATD) algorithms operate on a set of data statistics or features rather than on the raw radar sensor data. These features are selected based on their ability to separate target data samples from background clutter samples. The ATD algorithms often operate on the features through a set of parameters that must be determined from a set of training data that are statistically similar to the data set to be encountered in practice. The designer usually attempts to minimize the number of features used by the algorithm -- a process commonly referred to as pruning. This not only reduces the computational demands of the algorithm, but it also prevents overspecialization to the samples from the training data set. Thus, the algorithm will perform better on a set of test data samples it has not encountered during training. The Optimal Brain Surgeon (OBS) and Divergence Method provide two different approaches to pruning. We apply the two methods to a set of radar data features to determine a new, reduced set of features. We then evaluate the resulting feature sets and discuss the differences between the two methods.

The quadratic polynomial detector (QPD) and the radial basis function (RBF) family of detectors -- including the Bayesian neural network (BNN) -- might well be considered workhorses within the field of automatic target detection (ATD). The QPD works reasonably well when the data is unimodal, and it also achieves the best possible performance if the underlying data follow a Gaussian distribution. The BNN, on the other hand, has been applied successfully in cases where the underlying data are assumed to follow a multimodal distribution. We compare the performance of a BNN detector and a QPD for various scenarios synthesized from a set of Gaussian probability density functions (pdfs). This data synthesis allows us to control parameters such as modality and correlation, which, in turn, enables us to create data sets that can probe the weaknesses of the detectors. We present results for different data scenarios and different detector architectures.

This paper develops a method to model the two parameter Constant False Alarm Rate (CFAR) detector as an intensity test. For optimality we show that the shape of the stencil should be matched to the radial intensity profile of the target. Using principal component analysis (PCA) we show experimentally that the first gamma kernel is a good approximation to the target profile, which explains the good results of the (gamma) -CFAR detector, and may lead to configurable stencils for better detectability.

This paper explores statistical pose estimation in SAR ATR using a recently proposed training method based on information theory. The theory of training with information theoretic learning is briefly summarized. Different pose estimator topologies and training criteria are employed. Experimental results in the MSTAR I and II show that our proposed method is capable of producing 1-DOF and 2-DOF pose estimations and we show the dependence of the training parameters on performance.

Variability within a class of target vehicles seriously impacts the performance of a target recognition system. In this paper we present two main ideas about handling the intra- class variability. First, we develop a metric to quantify and understand the extent of variability within a class and second, we examine the class of T-72 tanks from the MSTAR public release data sets and attempt to make templates representative of the whole class.

Correlation filters are attractive for synthetic aperture radar (SAR) automatic target recognition (ATR) because of their shift invariance and potential for distortion-tolerant pattern recognition. In particular, the maximum average correlation height (MACH) filter exhibits better distortion tolerance than other linear correlation filters. Despite its attractive features, it has been shown that the MACH filter relies perhaps too heavily on the average training image leading to poor clutter rejection performance. To improve the clutter rejection performance, we have introduced the extended MACH (EMACH) filter. We have shown that this new filter is better at rejecting clutter images while retaining the distortion tolerance feature of the original MACH filter. In this paper, we introduce a method to decompose the EMACH filter to further improve its performance. The paper describes the theory of this method and shows its potential advantages. Test results of this method using the public domain MSTAR data base are shown.

In this paper, we discuss the performance of correlation filter algorithms trained on Xpatch (synthetic) model images. In particular, we assess the performance of the maximum average correlation height (MACH) filter and distance classifier correlation filter (DCCF) correlation algorithms on a 3-class subset of the public release MSTAR data set. The successful performance of these algorithms on a 10-class problem has been reported in previous publications. The results reported to date however were based on filters trained on actual sensor data. The approach proposed here is viewed as a means to combine advantages of purely model-based techniques and the statistical/correlation based approaches. The paper reviews the theory of the algorithm, key practical advantages and details of test results on the 3-class public MSTAR database.

This paper is concerned with a representation of a target's complex Synthetic Aperture Radar signature that could be used for classification purposes. In this representation, the complex SAR signature of a desired target area (chip) as a function of the radar frequency and aspect angle ((omega) , (phi) ) i.e., the target area phase history data) are shown to directly map into the two-dimensional spectrum of the target's image via a nonlinear transformation; the same information base in the ((omega) , (phi) ) domain is shown to be retrievable from the digitally-spotlighted complex SAR signature of the desired target region. For the classification problem (Automatic Target Recognition), the resultant complex SAR signature in the ((omega) , (phi) ) domain is compared with a finite and discrete set of reference complex SAR signatures via a process that we refer to as signal subspace matched filtering.

One common problem in automatic target recognition (ATR) is the insufficient size of the training set. Methods have been proposed to counter act this shortcoming, such as the noisy interpolation theory, hints, new distance measure (tangent distance), virtual examples, etc. This paper presents the idea of creating virtual negative examples as severe distortions of the known class patterns. Two classifiers are studied, a perceptron and a Support Vector Machine (SVM) trained to recognize objects in synthetic aperture radar (SAR) images. They utilize the training set (positive examples) to create the discriminant function of each class in the conventional way. On the other hand, the virtual negative examples will help determine the regions where the discriminant function should yield a low value. The experimental results show that incorporating the negative examples improves greatly (nearly 50 percents improvement) the confuser rejection rates.

This paper presents an approach for recognizing occluded vehicle targets in Synthetic Aperture Radar (SAR) images. Using quasi-invariant local features, SAR scattering center locations and magnitudes, a recognition algorithm is presented that successfully recognizes highly occluded versions of actual vehicles from the MSTAR public data. Extensive experimental results are presented to show the effect of occlusion on recognition performance in terms of Probability of Correct Identification, Receiver Operating Characteristic (ROC) curves and confusion matrices. The effect of occlusion on performance of this recognition algorithm is accurately predicted. Combined effects such as occlusion and measured positional noise, as well as occlusion and other observed extended operating conditions (e.g., articulation) are also addressed. Although excellent forced recognition results can be achieved at very high (70%) occlusion, practical limitations are found due to the similarity of unoccluded confuser vehicles to highly occluded targets.

Model-based HRR signature prediction and matching is a difficult problem due to target variations, lack of model fidelity, and sensor variations. For this reason, model-based HRR signatures often mismatch actual measured signatures. This paper introduces a novel method for predicting signatures by creating a Parametric HRR signature model from existing measured or model-based signatures. The method identifies potential three-dimensional persistent virtual scatterers and estimates their scattering patterns. The result is a parametric signature function of azimuth which better matches available signature data. The model effectively smoothes the signatures along the scatterer tracks. When used with synthetic data the technique helps eliminate model inaccuracies and uncertainties which manifest themselves in scatterer interference predictions. With sparse measured HRR profile observations, the parametric method smoothly interpolates profiles over azimuth. The parameterization method is aided by a Modified Inverse Radon Transform for persistent scatterer localization and a Fourier decomposition for scattering pattern approximation. Results using the predicted synthetic signatures in ATR multi-class problem are presented and compared with the performance based on the original synthetic and measured signatures. Results from these experiments show that this method provides a modest increase in ATR performance for synthetic signatures. The performance with sparse measured data is greatly improved and approaches dense measured data performance.

This paper presents the estimated 'electrical lengths' of ten ground targets from measured high range resolution (HRR) data. The HRR data is derived from stationary synthetic aperture radar (SAR) data through reversing the process used to form the SAR images. These estimated ground target 'electrical lengths' are used in a length-only, forced decision automatic target recognition (ATR) algorithm to determine the impact of using the target's 'electrical length' as a single feature. Various levels of noise are also folded into the HRR profiles to study the impact of the signal-to-noise ratio (SNR) on the length-only ATR performance.

In contrast to Synthetic Aperture Radar (SAR), High Range Resolution (HRR) radar may economically provide satisfactory target resolution when applied to moving targets scenarios. We have devised a series of new physics-based HRR moving target models with different degrees of simplification. These models represent the scatterers from both targets and clutter equally. By employing these models, we can unify the studies of both clutter suppression and target feature extraction into a single topic of model parameter estimation. Therefore, finding reliable parameter estimation algorithms based on these models becomes an important topic for target identification using HRR signatures. This paper derives and presents two feasible parameter estimation algorithms. The first algorithm (1DPE) reduces the 2D-estimation problem to two 1D-estimation problems, and solves the problems by employing some mature 1D-estimation algorithms. The second algorithm (2DFT) utilizes the 2D Discrete Fourier Transform (DFT) to estimate the model parameters by simply applying the 2D DFT to the HRR data, and obtaining the estimation of model parameters from the peaks of the 2D DFT. In order to verify the performance of these algorithms, we performed a series of simulation experiments and the experimental results are presented in this paper. Finally, a brief comparison of these two algorithms is also presented.

The problem of interfering scatters in high range resolution (HRR) radar data is addressed in this paper. We derive, using a scattering center representation of target, classifiers that can handle the unknown phases of the centers. We also show how to incorporate uncertainties in the magnitudes and positions of the scattering centers. The automatic target recognition (ATR) problem is discussed in a Bayesian setting, and we show how the uncertainties can be handled by such scattering center classifiers. Monte Carlo simulations are used to evaluate the performance and robustness of the classifiers for simple test cases, and data from electromagnetic prediction codes are used to illustrate the behavior on real targets.

This study summarizes recent algorithmic enhancements made to the AFRL/SNAA Systems-Oriented High Range Resolution (HRR) Automatic Recognition Program (SHARP) in the areas of multiple-look updating and sensor fusion. The benefits in improved 1-D Automatic Target Recognition (ATR) performance resulting from these enhancements are quantified. The study incorporates a unique method of estimating Bayesian probabilities by exploiting the fact that 1-D range profiles formed from Moving and Stationary Target Acquisition and Recognition (MSTAR) target chips overlap in azimuth. Thus, multiple samples of range profiles exist for the same target at very similar viewing aspects, but from independent passes of the sensor. ATR performance using the Bayesian technique is characterized first for an updating architecture that fuses probabilities over a fixed number of looks and then makes a 'classify or reject' decision. A second proposed architecture that makes a 'classify, reject, or take another measurement' decision is also analyzed. For both postulated architectures, ATR performance enhancement over the SHARP baseline updating procedure is quantified.

We study the moving target detection and feature extraction in the presence of ground clutter for airborne High Range Resolution (HRR) phased array radar. To avoid the range migration problems that occur in HRR radar data, we first divide the HRR range profiles into Low Range Resolution (LRR) segments. Since each LRR segment contains a sequence of HRR range bins, no information is lost due to the division and hence no loss of resolution occurs. We show how to use a Vector Auto-Regressive (VAR) filtering technique to suppress the ground clutter. Then we show how to estimate the target parameters of interest, including the target Doppler frequency, Direction-of-arrival (DOA), and complex amplitude and range frequency of each target scatterer. A moving target detector based on a Generalized Likelihood Ratio Test (GLRT) detection strategy is also derived. Numerical results are provided to demonstrate the performance of the proposed algorithms.

A classifier based on a syntactic approach is developed for High range resolution (HRR) radar target recognition. An attribute grammar is used to represent the structure of an HRR signature and an error-correcting parsing mechanism is implemented to extract peaks in the HRR profile and suppress the extraneous spikes. In the training phase, an error correcting grammatical inference technique is employed for structural inference of HRR signatures using a positive sample set. Recognition is done using a minimum distance classifier where Levenshtein error measure is used as the distance metric. The error-correcting parsing procedure for peak extraction is used to perform both inference and recognition. Experiments performed using public release MSTAR database indicate that this approach has sufficient discrimination power to perform target detection in HRR signatures.

This paper presents recent ATR results with High Range Resolution (HRR) profiles used for classification of ground targets. Our previous work has demonstrated that effective HRR-ATR performance can be achieved if the templates are formed via Singular Value Decomposition (SVD) of detected HRR profiles and the classification is performed using normalized Matched Filtering (MF) [1, 2]. It had been shown theoretically in [1, 2] that the eigen-vectors are the optimal feature set representation of a collection of HRR profile vectors and we had proposed to use the dominant range- space eigen-vectors as templates, known as Eigen-Templates (ET). However, in [1, 2], HRR-ATR performance using the Eigen Template-Matched Filter (ETMF) combination had been applied to the forced decision case only using the XPATCH data sets. In this paper, we demonstrate the effectiveness in HRR- ATR performance of the ETMF approach by incorporating unknown target scenario [4]. All results in this paper use the public release MSTAR data. Furthermore, in our earlier work, HRR testing data was used without any additive noise, where it was found that detected-HRR data preprocessed by Power Transform (PT) can enhance ATR performance. However, results and analysis presented in this paper demonstrate that PT pre- processing when applied to noisy observation profiles tend to obscure the target information in the HRR profiles considerably which in turn leads to considerable deterioration in HRR-ATR performance. Hence, we argue in this paper that PT pre-processing should be avoided in practice in all HRR-ATR implementations. Instead, we show that the proposed ETMF with appropriate alignment and normalization of template and observation profiles can achieve excellent HRR-ATR performance.Extensive simulation studies have been carried out to validate the proposed approach. Results are presented for different noise levels in terms of Receiver Operating Characteristics (ROC) curves.

The goal of this paper is to demonstrate the benefits of a tracking and identification algorithm that uses a belief data association filter for target recognition. By associating track and ID information, the belief filter accumulates evidence for classifying High-Range Resolution (HRR) radar signatures from a moving target. A track history can be utilized to reduce the search space of targets for a given pose range. The technique follows the work of Mitchell and Westerkamp by processing HRR amplitude and location feature sets. The new aspect of the work is the identification of multiple moving targets of the same type. The conclusions from the work is that moving ATR from HRR signatures necessitates a track history for robust target ID.

This paper is an initial exploration into the effects of range resolution on Automatic Target Recognition (ATR) algorithms based on High Range Resolution (HRR) signatures. The theoretical performance of a two-class, forced-decision classifier is used to quantify the effects of radar resolution on ATR performance. The classifier employed in this study is a forced-decision instantiation of the matched subspace classifier (MSC) developed under the DARPA TRUMPETS program. The paper also examines effects of range resolution on the separability of individual HRR profiles. This work is supported by DARPA/SPO under the MSTAR Enhancements (HBTI) program and in cooperation with AFRL/SNAA.

Complex spatial filters derived from maximizing a unique objective function are applied to Automatic Target Recognition (ATR) in Synthetic Aperture Radar (SAR) both to grayscale images and directly to phase history (sensor) data. For a limited class of MSTAR targets, a high degree of discrimination is shown in both cases. Applications to tracking are discussed and some areas for future investigations are indicated.

We present a mathematical framework for investigating the spatial and temporal covariance properties of wideband radar signatures from complex targets. The phrase 'complex target' is used to describe the target consisting of a large number of discrete scattering centers distributed over an electrically large volume. The spatial covariance properties of the scattered field are found to depend primarily on the volume density of the scattering centers and the scattering geometry. The temporal covariance properties are found to depend on the motion of the individual component scatterers of the target. The results are presented within a general framework that includes bistatic scattering geometries. Characterizing signature covariance properties is important because they play a significant role in determining the performance of automatic target recognition systems as well as in the design of optimal detection and classification algorithms.

Geometric invariants which completely characterize the size and shape of three dimensional configurations of scattering centers on moving targets can be extracted from signals from ranging sensors such as radar. These invariants can potentially be used to fingerprint and track specific moving targets even when information about changes in the targets orientation between observations is unreliable and when no prior model exists for the target. This technology also creates new possibilities for more robust target recognition.

Recent research in invariant theory has determined the fundamental geometric relation between objects and their corresponding 'images.' This relation is independent of the sensor (ex. RADAR) parameters and the transformations of the object. This relationship can be used to extract 3-D models from image sequences. This capability is extremely useful for target recognition, image sequence compression, understanding, indexing, interpolating, and other applications. Object/image relations have been discovered for different sensors by different researchers. This paper presents an intuitive form of the object/image relations for RADAR systems with the goal of enhancing interpretation. This paper presents a high level example of how a 3-D model is constructed directly from RADAR (or SAR) sequences (with or without independent motion). the primary focus is to provide a basic understanding of how this result can be exploited to advance research in many applications.

We consider a method for estimating classification performance of a model-based synthetic aperture radar (SAR) automatic target recognition system. Target classification is performed by comparing an unordered feature set extracted from a measured SAR image chip with an unordered feature set predicted from a hypothesized target class and pose. A Bayes likelihood metric that incorporates uncertainty in both the predicted and extracted feature vectors is used to compute the match score. Evaluation of the match likelihoods requires a correspondence between the unordered predicted and extracted feature sets. This is a bipartite graph matching problem with insertions and deletions; we show that the optimal match can be found in polynomial time. We extend the results in 1 to estimate classification performance for a ten-class SAR ATR problem. We consider a synthetic classification problem to validate the classifier and to address resolution and robustness questions in the likelihood scoring method. Specifically, we consider performance versus SAR resolution, performance degradation due to mismatch between the assumed and actual feature statistics, and performance impact of correlated feature attributes.

Performance prediction of SAR ATR has been a challenging problem. In our previous work, we developed a statistical framework for predicting bounds on fundamental performance of vote-based SAR ATR using scattering centers. This framework considered data distortion factors such as uncertainty, occlusion and clutter, in addition to model similarity. In this paper, we present an initial study on learning the statistical distributions of these factors. We focus on the development of a method for learning the distribution of a parameter that encodes the combined effect of the occlusion and similarity factors on performance. The impact of incorporating such a distribution on the accuracy of the predicted bounds is demonstrated by comparing bounds obtained using it with those obtained assuming simplified distributions. The data used in the experiments are obtained from the MSTAR public domain under different configurations and depression angles.

This paper continues the study reported in Ref. 1 and Ref. 2 trading off the fundamental ATR performance capability (i.e., algorithm-independent) of various SAR design options. The previous papers considered the performance impact of SAR range/cross-range resolution and compared the use of 1-D HRR (high-range-resolution radar) versus 2-D SAR, versus multisensor, 3-D SAR. The work reported here extends the SAR and HRR results of Ref. 2 to include aspect diversity in the SAR measurements. We show that SAR and HRR are benefited by multi-aspect measurements mostly because multiple views add diversity: poorer views benefit from having better views combined in a multi-aspect classifier. Finally, as a proof of concept, multi-aspect diversity is incorporated into an existing SAR ATR classifier; performance of an MSTAR 10-class MSE classifier is shown to improve substantially. A major tenet is verified by the experimental results: added measurement domains, such as aspect diversity, which separate the target signature vectors in the observation space, make it easier to obtain better target classification, enhanced false- alarm rejection, and robustness to unknown statistics.

The principal design challenge in Automatic Target Recognition (ATR) is signature variability. Target signatures are known to vary significantly as a function of numerous nuisance parameters, such as aspect angle, depression angle, component articulation, etc. There are several methods that have been used to manage nuisance parameters and the resulting signature variability. The focus of this paper is to define the various methods, provide examples, and illustrate the most common methods. A simple example will be used to compare the performance of the sampling, estimation, and integration methods.

The performance of an automatic target recognition (ATR) system for synthetic aperture radar (SAR) images is generally dependent upon a set of parameters which captures the assumptions made approximations made in the implementation of the system. This set of parameters implicitly or explicitly determines a level of database complexity for the system. A comprehensive analysis of the empirical tradeoffs between ATR performance and database complexity is presented for variations of several algorithms including a likelihood approach under a conditionally Gaussian model for pixel distribution, a mean squared error classifier on pixel dB values, and a mean squared error classifier on pixel quarter power values. These algorithms are applied under a common framework to identical training and testing sets of SAR images for a wide range of system parameters. Their performance is characterized both in terms of the percentage of correctly classified test images and the average squared Hilbert-Schmidt distance between the estimated and true target orientations across all test images. Performance boundary curves are presented and compared, and algorithm performance is detailed at key complexity values. For the range of complexity considered, it is shown that in terms of target orientation estimation the likelihood based approach under a conditionally Gaussian model yields superior performance for any given database complexity than any of the other approaches tested. It is also shown that some variant of each of the approaches tested delivers superior target classification performance over some range of complexity.

Synthetic Aperture Radar (SAR) image scene modeling tools are of high interest to Automatic Target Recognition (ATR) algorithm evaluation because they allow the testing of ATR's over a wider range of extended operating conditions (EOCs). Typical EOCs include target aspect, target configuration, target obscuration, and background terrain variations. A baseline terrain image synthesis technique empirically derived probability density functions (pdfs) for various terrain types from measured data to allow the simulation of user defined scenes. Initial full scene simulation experiments that applied this technique to the MSTAR data showed that using measured images as a data source for creating distribution functions in artificial scenes can introduce error, unless the proper spatial autocorrelation is also modeled. Measured SAR scene pixels have non-zero autocorrelation that blurs edges between different terrain types and creates a texture in the clutter regions of the image. Unfortunately, applying simple blurring techniques, such as a moving weighted window, to model autocorrelation mutes the second and higher moments of the pixel amplitude statistics. We propose a technique that models spatial autocorrelation while preserving the desired the amplitude statistics within each defined terrain class group.

A new very high-frequency compact radar range has been developed to measure scale models of tactical targets. This compact range has demonstrated very good signal-to-noise and is useful in measuring low observable targets. In addition to normal ISAR imaging of targets (range vs. horizontal cross- range), the system can also produce two-dimensional images in azimuth and elevation (vertical cross-range vs. horizontal cross-range). The 1.56 THz transceiver uses two high-stability optically pumped far-infrared lasers, microwave/laser side- band generation for frequency sweep, and a pair of Schottky diode receivers for coherent integration. Measurements made on 1/16th scale models of tactical targets, simulating W-band frequencies, allows the formation of images of very high cross-range resolution (3.5 cm full scale) while still integrating over a reasonably small angular extent (2.5 degrees). The results from several targets that have been recently measured will be presented.

Automatic target recognition (ATR) and feature-aided tracking (FAT) algorithms that use one-dimensional (1-D) high range resolution (HRR) profiles require unique or distinguishable target features. This paper explores the use of statistical measures to quantify the separability and stability of ground target features found in HRR profiles. Measures of stability, such as the mean and variance, can be used to determine the stability of a target feature as a function of the target aspect and elevation angle. Statistical measures of feature predictability and separability, such as the Fisher and Bhattacharyya measures, demonstrate the capability to adequately predict the desired target feature over a specified aspect angular region. These statistical measures for separability and stability are explained in detail and their usefulness is demonstrated with measured HRR data.

The research community is constantly searching for a rich high range resolution (HRR) radar dataset against moving targets. Due to the lack of HRR sensors, however, these data are unavailable and progress in tracking and identification of moving targets has been slow. This effort represents an attempt to synthesize HRR profiles for ground moving targets from readily available stationary ground target measured and synthetic data. The measured data are taken from the extensive Moving and Stationary Target Acquisition and Recognition (MSTAR) program database. The synthetic data are generated using Xpatch scattering center models for ground targets. In both cases, the target data are segmented from any background clutter and then processed on a pulse-by-pulse basis to inject linear constant velocity motion onto the target. The results are displayed in a range/Doppler image. Targets exhibit range walk and Doppler smear as would be expected. A keystone technique is used to correct the linear distortions to the phase history and refocus the target HRR chips for sensors requiring longer HRR dwell times.

SAR images of sites containing many buildings or other structures can be difficult to interpret visually. This is because SAR signatures generated by such sites are both complex and yet often sparser than the panchromatic signatures so familiar to the human brain. It is often difficult to visually associate a given SAR signature component with a specific structure. This paper presents ongoing work aimed at developing interpretation aids for such SAR images. Specifically, the paper focuses on recently developed methods for rapidly predicting SAR signatures of buildings and then overlaying them precisely on SAR imagery as visual interpretation guides. Using 3-dimensional building models, complex multibounce SAR signatures are simulated at interactive speeds. Signatures from dominant single point or single line scatterers are similarly formed. Tradeoffs between signature simulation speed and precise signature detail are discussed. Also discussed are the sensor modeling requirements and the precision georegistration methods needed to position the simulated signatures over actual SAR imagery. The paper includes a discussion of techniques used and examples of simulated signatures overlaid on imagery using these methods. Finally, application of the SAR signature overlays and precision registration methods to image fusion is discussed.

A key issue in the development and deployment of model-based automatic target recognition (ATR) systems is the generation of target models to populate the ATR database. Model generation is typically a formidable task, often requiring detailed descriptions of targets in the form of blueprints or CAD models. Recently, efforts to generate models from a single 1-D radar range profile or a single 2-D synthetic aperture radar (SAR) image have met with some success. However, the models generated from these data sets are of limited use to most ATR systems because they are not three-dimensional. We propose a method for generating a 3-D target model directly from multiple SAR images of a target obtained at arbitrary viewing angles. This 3-D model is a parameterized description of the target in terms of its component reflector primitives. We pose the model generation problem as a parametric estimation problem based on information extracted from the SAR images. We accomplish this parametric estimation in the context of data association using the expectation-maximization (EM) method. Although we develop our method in the context of a specific data extraction technique and target parameterization scheme, our underlying framework is general enough to accommodate different choices. We present results demonstrating the utility of our method.

Proper evaluation of a pattern recognition system in the lab is paramount to its success in the field. In most commercial pattern recognition applications, such as breast cancer detection, optical character recognition, and industrial quality assurance, the boundaries and expectation of the system are well defined. This is due, at least in part, to an excellent understanding of the problem and data space for these applications. For these functions, a method for rigorous evaluation is well understood. However, the size and complexity of the data and problem spaces for automatic target recognition (ATR) systems is enormous. The consequences are that a complete understanding of how an ATR system will perform in practice is extraordinarily difficult to estimate. Thus the act of evaluating an ATR system becomes as important as its design. This paper compiles and reports the techniques used to evaluate ATR system performance. It surveys the specific difficulties associated with ATR performance estimation as well as approaches used to mitigate these obstacles.

The TESAR [Tactical Endurance Synthetic Aperture Radar (SAR)] system uses four algorithms in its three-stage algorithmic approach to the detection and identification of targets in continuous real-time, 1-ft-resolution, strip SAR image data. The first stage employs a multitarget detector with a built-in natural/cultural false-alarm mitigator. The second stage provides target hypotheses for the candidate targets and refines their angular pose. The third stage, consisting of two template-based algorithms, produces final target-identification decisions. This paper reviews the end- to-end ATR performance achieved by the TESAR system in preparation for a 1998 field demonstration at Aberdeen Proving Ground, Aberdeen, MD. The discussion includes an overview of the algorithm suite, the system's unique capabilities, and its overall performance against eight ground targets.

With the large amount of image data that can be produced in real-time by new synthetic aperture radar (SAR) platforms, such as Global Hawk, compression techniques will be needed for both transmission and storage of this data. Also to keep image analysts (IA's) from being overwhelmed, high-speed automatic target cueing and/or recognition (ATC, ATR) systems will be needed to help exploit this large amount of data in real-time. Past SAR image compression studies have used subjective visual ratings and/or statistical measures such as mean-squared-error (MSE) to compare compression performance. Statistical metrics are much more appealing than unreproducible biased visual interpretations. However, the use of statistical metrics, such as MSE, has practical limitations on SAR imagery due to the high frequency speckle noise that is characteristic. In this case, the MSE metric is dominated by how well the noise speckle is preserved -- a statistic that is of no consequence. Since the large amount of data that dictates the need for compression also dictates the need for ATR, a meaningful statistic would be ATR performance. This ATR performance metric would emphasize how well pixels on target are preserved. Therefore, we have investigated ATR performance using a wavelet compression technique, since this technique has achieved very high compression on other types of imagery. We have used the Rice University Computational Mathematics Laboratory's wavelet compression algorithm in conjunction with a 'synthetic discriminant function' (SDF) based ATR algorithm. The SDF technique was developed at Carnegie Mellon University and successfully applied to SAR imagery by the Northrop Grumman Science & Technology Center. This combination allows ATR performance to be parameterized as a function of compression rate. The SAR data used for this research was taken from the public-released MSTAR target and clutter data set. We show results for both target detection and target identification versus false alarms for varying compression rates.

Conventional image compression methods compress all regions of an image with a roughly uniform compression ratio. This means that any regions of special interest are degraded on an equal basis as the remainder of the image. Content-Based Image Compression (CBIC) methods assign different compression rates to different regions of an image according to their priorities, or according to the relative importance of the regions for certain applications. For example, for visual perception, we can assign different compression rates to different objects so that after compression the objects of interest satisfy certain MSE (Mean Square Error) requirements regardless of the overall compression rate. For optimal ATR (Automatic Target Recognition) performance, the recognition error rate might be optimized instead of MSE so that target recognition performance will be guaranteed at some desired level, and held constant throughout the entire image. In this paper, we introduce a content-based image encoder based on the popular DCT and wavelet transforms. Instead of selecting the DCT/wavelet coefficients that minimize the MSE to achieve optimum visual effects, we propose an algorithm to preserve those coefficients that minimize the recognition error. For any ATR system that utilizes the resulting compressed images, the recognition error is bounded by the information-theoretic distances. We employ Chernoff distances to compute the cost function of the recognition error. Compared to image compression methods optimized for visual perception, our results show that this CBIC method for ATR is able to achieve significantly more uniform ATR performance by assigning different compression rates to different regions.

Evaluation is a cornerstone of AFRL/SN research into target detection, tracking, and identification. A formal evaluation procedure provides quantitative feedback on algorithm performance and a means for monitoring algorithm improvement and for comparison of different algorithms on an objective and equal basis. AFRL/SN has invested in an Advanced Real-Time in the Cockpit Cell (ARC) facility as a part of the Collaborative Engineering Environment (CEE) that allows for rapid evaluation of target detection, tracking, and identification algorithms within a comprehensive simulation environment that includes target simulators, sensor simulators, analyst workstations, strike simulators and weapon flyout. A testbed is under development to support the evaluation of RF moving target technology as well as maturity assessments of the technologies for transition.

Detection, location and SAR imaging of moving targets in clutter have attracted much attention. Locations of moving targets in the SAR image are determined not only by their geometric locations but also by their velocities that cause their SAR images de-focused, smeared, and mis-located in the azimuth dimension. Furthermore, the clutters may cause the detection of moving targets more difficult. Several antenna array based algorithms have been proposed to re-locate the moving targets in the SAR image. With a linear antenna array, the clutters may be suppressed using multiple phase centers. However, there are only two parameters involved in a linear antenna array, i.e., number of receiving antennas and the distance between two adjacent antennas. These two parameters physically limit the capability to detect the accurate locations of fast moving targets and such as vehicles, and only slowly moving targets, such as walking people, can be correctly re-located. In this paper, we propose an antenna array approach where transmitting single wavelength signals are generalized to transmitting multiple wavelength signals (called multi-frequency antenna array SAR). We show that, using multi-frequency antenna array SAR, not only the clutters can be suppressed but also locations of both slow and fast moving targets can be accurately estimated. For example, using two-frequency antenna array SAR system with wavelengths (lambda) ₁ equals 0.03 m and (lambda) ₂ equals 0.05 m, the maximal moving target velocity in the range direction is 1 5 m/s while using single frequency antenna array SAR system with wave length (lambda) ₁ equals 0.03 m or (lambda) ₂ equals 0.05 m, the maximal moving target velocity in the range direction are 3 m/s or 5 m/s, respectively. A robust Chinese Remainder Theorem (CRT) is developed and used for the location of fast and slowly moving targets. Simulations of SAR imaging of ground moving targets are presented to illustrate the effectiveness of the multi-frequency antenna array SAR imaging algorithm.

In this paper we present a method of fusing evidence of targets in an observed SAR image. We have selected three simple feature types to perform these initial experiments. Fusion is performed by straightforward addition of log likelihoods over feature match types. In all cases among the three feature types tested it is observed that the probability of identification improves as feature types are added. The method of recognition is based on the probabilistic distance transform (PDT). This approach derives from traditional distance transform (DT) methods of matching target predictions (based either on training or model based predictions) to observed features. The PDT method retains the basic DT matching structure, including the advantages of fast processing and non-unique correspondences between predicted and observed features, while interpreting 'distance' in terms of spatial probability densities of predicted and observed features. The PDT matching approach then results in a statistic that can be treated as a likelihood of match between an observed set of features and a predicted target signature.

In Automatic Target Recognition (ATR) systems Receiver Operating Characteristics (ROC) curves are used to describe operating characteristics for changing threshold values of two aspects of importance to the users. the two aspects are the ability of the Automatic Target Recognition (ATR) system to detect targets and its ability to reject non-targets that is, not declare false alarms. These two abilities are represented in ATR systems analysis by the probability of detection and the false alarm rate. The problem of characterizing the confidence area for parameters of the probability distribution of ROC points is addressed and applied to the Moving and Stationary Target Acquisition and Recognition (MSTAR) DEMO III data using MATLAB.