Statistical Performance of classifiers for a maritime ATR Task

Abstract

This research explores the statistical performance of several classifiers (Bayes, nearest neighbor, and a neural network) on a maritime ATR problem. The features employed were derived from range profiles and inspired by the physical structure of the ship targets to maximize the generalizability of the classifiers. The ship targets were created using Pro Engineer (parametric technology corporation), facetized, and input into XPATCH. XPATCH was used to create range profiles from 0 to 30 degree aspect. A likelihood based confidence measure was employed to force the classifiers to output at 98% confidence. The confidence measure was based on a discriminant that was the distance between a classifier output and a template.