Experimental methods for estimating detection efficiencies of human and computational observers viewing low-contrast circular targets in acoustic noise are described. Sonographic images were simulated with signal and noise properties specified exactly. These images were presented to observers in two-alternative forced-choice (2AFC) experiments. Relative to the ideal observer of these images, i.e., the prewhitening matched filter, human observers were 60% efficient for detecting targets over a broad range of target energies and for both target polarities. Studies were limited to situations where target diameters were much larger than the correlation length of the noise. In that case, observers unable to decorrelate the noise showed no reduction in detectability as predicted by theory. For example, the efficiency of one computational observer, a nonprewhitening matched filter, was nearly ideal. Its response was proportional to that of the average human observer, which suggests a role for computational observers in image evaluation.
No takes yet. Share an insight, caveat, or question.
Insana et al. (1994) studied this question.