A Software Tool for ATR performance Characterization

Autonomous surveillance and weapon applications require the ability to perform ATR on data obtained from a number of sources (including infrared, synthetic aperture and inverse synthetic aperture). The algorithms generally employ tuning parameters and it is not evident given the non-linear and feedback aspects associated with ATR suites that optimization of individual algorithms within an ATR suite is equivalent to optimization of the ATR itself. The objective of the Phase I effort is to demonstrate the feasibility of developing a set of software tools for characterizing performance of ATR systems relative to a given data set. These tools will by used for the design, evaluation and optimization of ATR systems. The innovations in the proposed approach are: a) obtaining tighter error bounds with reduced biases and variances by combining multiple classifiers, b) exploring new results in neural networks for approximating posterior probabilities, and c) exploiting knowledge of the ATR system components to develop more accurate statistical models of performance. We propose to: 1) develop and implement a prototype of the proposed ATR performance characterization tools, 2) test the developed tools using synthetic and real ATR data, 3) study and document the characteristics and reliability of the developed performance tools. The SSC project team will be supported by SAIC (Dr. Arnold Williams) and Prof. Haralick, both of whom have done pioneering work in the area of ATR performance evalustion.