Abstract
Thorpe et al. (1996) showed that humans can detect the presence of object categories in natural images in as little as 150ms. The purpose of our study is to see if this rapid category detection is possible in the context of visually degraded images. Previous studies have demonstrated that humans can accurately process images in severely degraded conditions. Harmon and Julesz (1973), and Bachman (1991) demonstrate that only 18×18 pixels per face are sufficient for robust recognition, and these findings have been extended to the domains of objects and scenes by Torralba et al. (2007). Along the dimension of luminance depth, Mooney faces are a classic demonstration of visual processing working in extreme cases of luminance depth degradation. However, it is not clear whether or not rapid category detection is possible in visually degraded images.
In this study, we explore the phenomenon of rapid object categorization in the context of both singly and multiply degraded images. Our experimental setup is modeled after Kirchner et al.'s (2006) 2AFC Animal/Non-Animal detection task. In each trial, subjects are flashed a pair of natural scenes for 30ms where only one scene contains an animal. Subjects are directed to choose the image containing an animal and measured for accuracy in different conditions. In our first experiment, each condition contains images degraded to different degrees along one of the following dimensions: spatial resolution, luminance depth, contrast, inversion and reverse contrast. In our second experiment, each condition contains images degraded along pairs of the aforementioned dimensions. We find that even with severe degradation, such as natural images with only 50×38 pixels or 1 bit plane, humans can rapidly identify images containing animals with high accuracy. For multiply degraded images, we find subjects' performance impacted by a non-linear interaction between different types of degradation.
National Science Foundation Graduate Fellowship and the Arthur J. Chick Professorship.