Item

Abstract

What information do humans use during the identification of complex natural scenes? To address this question, we employed a technique described originally by Gosselin and Schyns (2001 Vision Research 41 2261 - 2271), which identifies image regions diagnostic for visual-recognition tasks. We applied this method to a task where subjects had to discriminate natural images. On each trial, one of four images was shown behind an occluding mask punctured by multiple randomly located Gaussian windows. Diagnostic image regions were computed by comparing masks that resulted in correct performance with masks leading to incorrect performance. During different sessions, we used either a constant-stimuli protocol with a fixed number of windows, or a staircase protocol to adjust the number of windows as a function of behavioural performance. In general, depending on the particular natural image, different regions were revealed as diagnostic. Results for the constant-stimuli and the staircase protocols were in good agreement. For the constant-stimuli protocol, we found that subjects' behavioural performance improved with training for some natural images. Diagnostic information generally did not show dramatic changes, although sometimes particular image regions became diagnostic with learning. These results demonstrate that reverse correlation can be used to reveal diagnostic regions in complex natural images.