Performance of a target identification algorithm as a function of the discriminant post-processing techniques utilized

Abstract

A target identification algorithm can be thought of as being comprised of a pre-processor, discriminant generator, and a post-processor. The pre-processor forms and otherwise conditions the target signatures. The discriminant generator forms scalar quantities that represent the closeness-of-fit of each signature to the target classes of interest. Finally, the post-processor utilizes those scalars to form a decision as to the target that the signature came from; i.e., the identity of the target being examined. In this paper, we start with a full set of discriminants generated by a particular pre-processor and discriminant generator operating on high-range-resolution (HRR) signatures of aircraft, and we perform various experiments to determine the effect on algorithm performance of applying various post-processing techniques. The overall target identification algorithm is described, numerous post-processing techniques are introduced, and their effects on performance are tabulated. It is shown that optimal combined performance of these techniques does not necessarily follow from combining the individual best-performing techniques. That is, an optimal post-processing architecture cannot be derived from a simple search of the diagonal of the multi-dimensional set of post-processing options.