Statistical inference for automatic target recognition systems

Traditionally, correlation-based target-recognition systems are evaluated by observation of the number of images successfully recognized and the errors made in the classification process. However, the rigorous use of hypothesis tests and confidence intervals in these evaluations does not appear with any regularity in the optics literature. The optical target-recognition community should adopt these standard methods to evaluate objectively the statistical performance of automatic target recognition systems. We review the necessary steps for making statistically significant inferences about the performance of automatic target recognition systems and show their application by reevaluating some previous results as case studies.