In this paper, we investigate the effect of increasingly sparse training data sets on target classification performance using
a template-based classifier. An often used method of template creation employs averaging of multiple target training
chips for a predefined coverage swath. The inclusion of too many training chips results in a blurring of the predominant
scatterers while averaging of too few training chips results in poor edge resolution. We use the public MSTAR data set
to show that using all appropriate images for each template may not result in the best ATR performance. We
successfully demonstrate the ability to reduce training data collection requirements by requiring fewer training chips per
Performance of Automatic Target Recognition (ATR) algorithms for Synthetic Aperture Radar (SAR) systems relies
heavily on the system performance and specifications of the SAR sensor. A representative multi-stage SAR ATR
algorithm [1, 2] is analyzed across imagery containing phase errors in the down-range direction induced during the
transmission of the radar's waveform. The degradation induced on the SAR imagery by the phase errors is
measured in terms of peak phase error, Root-Mean-Square (RMS) phase error, and multiplicative noise. The ATR
algorithm consists of three stages: a two-parameter CFAR, a discrimination stage to reduce false alarms, and a
classification stage to identify targets in the scene. The end-to-end performance of the ATR algorithm is quantified
as a function of the multiplicative noise present in the SAR imagery through Receiver Operating Characteristic
(ROC) curves. Results indicate that the performance of the ATR algorithm presented is robust over a 3dB change in
Template-based classification algorithms used with synthetic aperture radar (SAR) automatic target recognition (ATR)
degrade in performance when used with spatially mismatched imagery. The degradation, caused by a spatial mismatch
between the template and image, is analyzed to show acceptable tolerances for SAR systems. The mismatch between
the image and template is achieved by resampling the test imagery to different pixel spacings. A consistent SAR dataset
is used to examine pixel spacings between 0.1069 and 0.2539 meters with a nominal spacing of 0.2021 meters.
Performance degradation is observed as the pixel spacing is adjusted, Small amounts of variation in the pixel spacing
cause little change in performance and allow design engineers to set reliable tolerances. Alternatively, the results show
that using templates and images collected from slightly different sensor platforms is a very real possibility with the
ability to predict the classification performance.
A multi-stage Synthetic Aperture Radar (SAR) Automatic Target Recognition (ATR) system is analyzed across images
of various pixel areas achieved by both square and non-square resolution. Non-square resolution offers the ability to
achieve finer resolution in the range or cross-range direction with a corresponding degradation of resolution in the cross-range
or range direction, respectively. The algorithms examined include a standard 2-parameter Constant False Alarm
Rate (CFAR) detection stage, a discrimination stage, and a template-based classification stage. Performance for each
stage with respect to both pixel area and square versus non-square resolution is shown via cascaded Receiver Operating
Characteristic (ROC) curves. The results indicate that, for fixed pixel areas, non-square resolution imagery can achieve
statistically similar performance to square pixel resolution imagery in a multi-stage SAR ATR system.
Automatic target recognition (ATR) performance is a function of image quality and its representation in the signature model generation and used in the ATR training process. This paper reports ATR performance as a function of synthetic aperture radar (SAR) image quality parameters including clutter-to-noise ratio (CNR) and multiplicative noise ratio (MNR). Images with specified image quality values were produced by introducing controlled degradations to the MSTAR public release data. Two different families of ATR algorithms, the statistical model-based classifier of DeVore, et al., and optimal tradeoff synthetic discriminant function (OTSDF) are applied to those data. Target classification accuracy was measured as a function of CNR/MNR for both the training and test data, indicating sensitivity of performance to a priori knowledge of these particular image quality parameters. Confusion matrices are expanded to include target aspect bins, providing visibility into performance as a function of aspect angle.
Many applications which process radar data, including automatic target recognition and synthetic aperture radar image formation, are based on probabilistic models for the raw or processed data. Often, data collected from distinct directions are assumed to represent independent observations. This assumption is not valid for all data collection scenarios. A range of models can be developed that allow for successively more complex dependencies between measured data, up to deterministic computational electromagnetic models, in which observations from different orientations have a known relationship. We consider models for the autocovariance functions of nonstationary processes defined on a circular domain that fall between these two extremes. We adopt a model of covariance as a linear combination of periodic basis functions and address maximum-likelihood estimation of the coefficients by the method of expectation-maximization (EM). Finally, we apply these estimation methods to SAR image data and demonstrate the results as they apply to target recognition.
Model-based approaches to automatic target recognition (ATR) generally infer the class and pose of objects in imagery by exploiting theoretical models of the formed images. Recently, we have performed an evaluation of several statistical models for synthetic aperture radar (SAR) and have conducted experiments with ATR algorithms derived from these models. In particular, a one-parameter complex Gaussian model, classically used to model diffuse scattering, was shown to deliver higher recognition rates than a one-parameter quarter-power normal model on actual SAR data. However an extended, two-parameter quarter-power model was consistently a better fit to the data than a corresponding two-parameter Gaussian model. In this paper, we apply Rician, gamma, and K distribution models, which are two-parameter extensions of the complex Gaussian and quarter-power normal models, to ATR from SAR magnitude imagery. We consider maximum-likelihood estimation of unknown model parameters and apply the resulting training and testing algorithms to actual SAR data. We show that the K distribution model performs better than the Rician and gamma models for both large and small sample sizes. The one-parameter complex Gaussian model performed slightly better than the K model overall. For small sample sizes, this is likely due to the relative stability in estimating only one model parameter. For large sample sizes this is likely due to a lack of persistence in specular reflections over the large angular intervals required to obtain large samples.