Abstract
Purpose: To characterize the linear and nonlinear mechanisms underlying feature detection in human vision. Methods: Observers detected the presence or absence of a vertical, target bar flashed for 30 ms, embedded in spatiotemporal noise (vertical bars with random, uniformly distributed luminances). The noise level (luminance range) was set so that observers would be at threshold performance. We estimated two spatiotemporal sensory filters underlying observers' performance, using psychophysical reverse correlation (Beard & Ahumada, Perception 26:38, 1997): in this technique, the individual noise samples are averaged separately for trials corresponding to hits, misses, false alarms, and correct rejections. The first filter was computed by averaging the noise samples themselves, thus obtaining the first-order kernel of the detection mechanism. The second filter was computed from the variance of the noise, thus obtaining the second-order kernel. Results: The first-order (linear) kernel peaked at target spatiotemporal location (i.e. it resembled the target), with spatial inhibitory flanks. The second-order (nonlinear) kernel, on the other hand, peaked before target presentation. Conclusions: Because the second-order kernel peaks before the first-order kernel, detection appears to involve two stages: an energy extraction stage (possibly supported by nonlinear processing of complex cells) followed by a feature identification stage (possibly supported by linear processing of simple cells). These results are consistent with energy models of feature detection (Morrone & Burr, Proc. R. Soc. Lond. B 235: 221–245, 1988).