<title>Class separability assessments and MSE algorithm robustness</title>

Abstract

Operational Synthetic Aperture Radar (SAR) based automatic target recognition (ATR) systems will encounter a wide range of target types, some of which will be variants of pre- mission training targets. Previous work measured classification performance when training and testing on different vehicle variants and assessed intra-class separability based on an empirical estimate of the mean square error (MSE) probability density function. This research showed a significant degree of intra-class signature variability for selective targets, resulting in a difficult ATR problem. The benefits of using mixture templates were demonstrated with respect to classification performance as well as pose prediction. This paper extends this analysis by considering the signature variability attributed to extended operating conditions such as depression angle and articulation. Furthermore, it demonstrates improved performance robustness is possible using an MSE classifier with appropriate normalization and segmentation. Additionally, a simple technique for minimizing the impact of localized error sources on MSE algorithms is discussed. Finally, error surfaces associated with missed classifications are shown to b similar in both space and amplitude, suggesting finer target discrimination may require improved feature sets and or adaptive refinement algorithms for handling both deterministic and random error sources associated with the observation to template.