Multiple Hypothesis Tests For Robust Radar Target Recognition

In radar automatic target recognition (ATR), an input that does not originate from one of the known training classes may be forcibly declared as one, thereby reducing system reliability. Furthermore, the feature space may contain ambiguous regions wherein declaration for a single class is not possible. In this paper, classification is viewed as a multiple hypothesis problem facilitating the rejection of inputs from classes that were absent from the training data, or that originated from ambiguous regions of the feature space, alongside the traditional declarations for the trained classes. Use of the bootstrap, to estimate the p-values used in the multiple hypothesis tests, is advocated during implementation to prevent the need for assumptions on the underlying statistical characteristics of the data. Experimental through-the-wall radar imaging (TWRI) data are used to validate the proposed techniques, and the proposed classifier is deemed more reliable than a conventional minimum distance (MD) classifier due to its ability to reject data that cannot be correctly classified. High rates of correct-classification are still obtained for valid input data.