Abstract
Humans make eye movements because sensitivity across the retina varies greatly. We have derived the most efficient eye movement strategy for foveated visual systems, which are searching for targets embedded in spatial noise. The ideal searcher considers the information collected thus far, and always chooses the next fixation location in order to maximize the probability of finding the target after the data from that fixation is collected. The ideal performs a simultaneous optimal Bayesian estimation of both the signal and external noise at every potential target location on every fixation. The behavior of the ideal searcher depends on its sensitivity to the target across the visual field. The sensitivity can be characterized as a signal-to-noise ratio, and it can be modeled as having two components: one due to the external noise, and the other due to internal variability. We estimated the relative contributions of these two components in psychophysical experiments. Human observers were asked to detect a Gabor patch (6 cpd) embedded in 1/f noise, at 25 different retinal locations spread evenly across the visual field (0, 1.5, 3, and 4.5 deg eccentricity). At each retinal location, we measured psychometric functions at four external noise levels (0, 0.05, 0.1, and 0.2 rms contrast). The relationship between signal energy at the detection threshold and the variance of the external noise was found to be approximately linear at all eccentricities. This result generalizes to 1/f noise the commonly reported trends in foveal white-noise masking experiments. We use the results of our noise masking detection experiments to obtain an essentially parameter free model of optimal eye-movements during visual search. Among other things, this allows us to measure the absolute efficiency of human eye movement patterns. We are currently comparing the behavior of this optimal searcher to human behavior.