Improving high range resolution ATR performance with average absolute error metric

As a complementary mode to synthetic aperture radar (SAR), high range resolution (HRR) is mainly developed to identify moving ground targets automatically. Previous studies have demonstrated many promising solutions in HRR Automatic Target Recognition (ATR). Most of them popularly build on template-based approach and employ mean square error (MSE) metric in matching algorithm to measure the likelihood between test profiles and template profiles. However, discussion about whether MSE works as it is expected is inadequate. This paper represents a preliminary investigation into MSE metric used in matching algorithm. We firstly illustrate MSE metric theoretically, and then indicate its shortcomings such as random coefficients distorting the difference between profiles in mathematical calculation, excessive emphasis on large difference, and lack of reasonable interpretation in physical distance space. An alternative metric named average absolute error (AAE) is utilized to replace MSE. It is directly based on the distance between profiles and is able to provide a clear physical interpretation. The experiment results show that no matter estimated aspect angles are given or not during classification, AAE-based HRR ATR significantly outperforms the baseline MSE-based HRR ATR.