Predicting human observer performance with response equivalent delta-T

Abstract

Blackwell described observer performance in terms of three variables: target size, contrast, and overall brightness. However, Blackwell's experiments involved uniform targets and backgrounds. Formulations for effective contrast to enable applying his data to generalized problems have resulted in substantial differences between observed and expected results. Response equivalent delta-T (REDT) is designed to solve this problem. REDT is a contrast metric that is defined in terms of the distance between a signature point and an experimentally derived threshold response surface. It provides more realistic predictions of target detectability than more customary contrast metrics in many situations involving nonuniform targets and backgrounds. In an appropriately controlled experiment, a correlation coefficient of 0.982 is obtained between observer data and REDT values. Although REDT has not been demonstrated to significantly outperform other contrast metrics in realistic applications, the scatter in such data can be understood in terms of uncontrolled variations in clutter, target size, and second-order image statistics, which describe texture. Gray-level cooccurrence matrices provide a plausible tool for understanding situations in which REDT fails to perform well. REDT is most appropriate when targets are small, resolution is low, and clutter is moderate to high.