Innovations in Physical Modeling and Statistical Exploitation of Electromagnetic Target Signatures

ABSTRACT: Feature extraction and target recognition suffer from a lack of a reliable model for both exploitable target features and the electromagnetic signature they possess. Signature data are often hard to interpret and invert to recover the target robustly. Bayesian learning approaches to statistical pattern recognition are based on the use of training sets of inputs and outputs, a data model, and statistical priors on the weight parameters controlling the outputs, but the main drawback of this approach is its use of training sets whose predictive properties are limited by their quality and comprehensiveness. Rigorous physical models, derivable from first principles, for the complete forward problem must constrain the system output in a more realistic, predictable way. All training sets depend on noise and resolution levels which can potentially grossly amplify the errors in predictions or classifications made about an unknown. Physics must be employed to work symbiotically in constraining such errors and thus improve the confidence levels of statistical inferences drawn from conventionally acquired training sets. FiveFocal and the University of New Mexico"s approach has a number of innovative components: innovative physical models, statistical assessment tools, parametric target representation, and a comprehensive sensor simulation facility. BENEFIT: The developments in statistical signature exploitation, physical modeling, and system analysis have significant commercial opportunity beyond the immediate defense application addressed in this proposal as it addresses an approximately $300 M/year commercial market. A long term goal for the sensors, detectors, and optics industry is an end to end simulation and modeling capability that accurately predicts performance of an as-manufactured imaging system, e.g. a robotic vision system for assembly work, or a consumer cell phone camera. Consumer and industrial electronics industries will realize tremendous gains in product development efficiency if improved system modeling can reduce the number of prototype iterations.