Performance improvement in ATR from dimensionality reduction

A thresholding method for the reduction of dimensionality applied to test statistics of an M-ary composite hypothesis testing problem, with maximum likelihood (ML) estimates incorporated instead of true parameters, is developed. The ML estimates are obtained from training sets of small size. The thresholding method selects only the entries in the testing vector that contain a large amount of information for discriminating among M hypotheses. The information measure is a plug-in version of the relative entropy with one of two distributions known. The method is promising for the exponential family. The performance of the test with a reduced number of dimensions is analyzed by applying a theory of asymptotic expansions of integrals. The study is of interest in automatic target recognition (ATR)