An image metric-based ATR performance prediction testbed

Currently, automatic target recognition (ATR) evaluation techniques use simple models, such as quick-look models, or detailed exhaustive simulation. Simple models cannot accurately quantify performance, while the detailed simulation requires enumerating each operating condition. A need exists for ATR performance prediction based on more accurate models. We develop a predictor based on image measures quantifying the intrinsic ATR difficulty on an image. These measures include: CFAR, power spectrum signature, probability of edge etc. We propose a two-phase approach: a learning phase, where image measures are computed on set of test images, and the ATR performance measured; and a performance prediction phase. The learning phase produces a mapping, valid across various ATR algorithms, even applicable when no image truth is available (e.g., evaluation for denied area imagery). We present a performance predictor using a trained classifier ATR constructed using GENIE, a tool from Los Alamos.