Abstract
The visual perceptual qualities of contrast, orientation, spatial frequency and color are strongly and complexly affected by the relationship between any given target and its context. For example, the perception of surface brightness is determined by the luminance contrast, orientation, spatial frequency and spectral characteristics of a target and these same parameters at every point in the rest of the scene. It remains unclear how interactions among these basic stimulus features are represented in the visual brain, and why target and context features affect each other in such extraordinarily complicated ways.
Here we suggest that, as a means of contending efficiently with the full range of naturally occurring visual stimuli, the human visual system encodes the probability distributions of the co-occurrence of these basic visual features in typical scenes. An advantage of this strategy of vision is that it makes use of the full coding capacity of the system in any typical situation. Since specific values of stimulus qualities such as luminance contrast, orientation, spatial frequency and spectral distribution co-occur to different degrees in different natural contexts, the pertinent representations of any one of these characteristics should, in these terms, be based on the conditional probability distribution of the relevant characteristic at one location, given the probability of occurrence of the possible values of all these other features at both the same and other locations.
To test the merits of this concept of vision, we analyzed a database of natural images, thus obtaining the relevant probability distributions of co-occurring contrasts, orientations, spatial frequencies and spectral distributions. We show that these probability distributions can indeed account for the wide range of target-context effects apparent in the perception of contrast, orientation, spatial frequency and color, supporting a wholly statistical basis for these perceptual phenomena.
This work was supported by the NIH, the AFOSR, and the Geller Endowment