We compared human detection of visual targets in noisy images with that of a theoretically optimum matched filter. Using a small thin target with vertically aligned markers, we obtained hyperefficient detection as high as 91% as compared with the theoretical optimum, a value far exceeding the 30–50% value typically reported. When the markers were removed, detection efficiencies degraded to an average of 27%, even though subjects were aware that the target was always placed in the center of a reasonably small panel. Using a nine-alternative forced-choice experiment, we compared detection by human observers with a matched-filter computational observer on a trial-by-trial basis. With the markers present, when humans missed the correct panel, they most often chose the panel with the second-highest decision variable output from the computational observer, suggesting that the template-matching model is a good one. To model results without the markers, we included location uncertainty and additional noise sources in the template matching of the computational observer. A location uncertainty of only 1 pixel, corresponding to a retinal distance of ≈12 μm, a dimension of the order of the size of the receptive field of photoreceptors, explained the psychometric data. With the marker present, the model suggests that hyperefficient detection is obtained by limiting target location uncertainty to <6 μm. Together these results give important new insights into human visual detection mechanisms.
© 2001 Optical Society of AmericaFull Article | PDF Article