Framework for ATR Performance Prediction and Optimization

We propose to design a performance prediction/optimization framework for Automatic Target Recognition (ATR) algorithm suites. This framework will be applicable to individual algorithms, groups of algorithms, and whole ATR suites. For the ATR evaluator, it offers the ability to predict the potential for performance improvement in a given algorithm or suite, without having to fully implement it. For the ATR designer, it offers the ability to optimize free parameters of algorithms and provide guidance on where algorithms can be optimized individually and where nonlinear relations between algorithms require them to be optimized together. Our proposed solution is centered on a systematic analysis of current state-of-the-art ATR methods in order to determine commonalities between algorithms in terms of input types, output types, and performance criteria. A taxonomy of ATR methods is used as a knowledge base for modeling algorithm input and output types, and determining performance criteria. Once performance as a function of algorithm parameters has been determined, optimization is performed. A limited software prototype will be implemented and applied to one ATR suite for feasibility demonstration, to be followed up in Phase II with a full scale implementation and validation applicable to a range of algorithms.