Purchase this article with an account.
Jonathan Victor, Daniel Thengone, Mary Conte; Characterizing the salience and interactions of informative image statistics. Journal of Vision 2011;11(11):1164. https://doi.org/10.1167/11.11.1164.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Visual perception of natural scenes depends on the processing of image statistics. They can drive segmentation of an image into objects, indicate the material composition of a surface, and provide the gist of a scene. To study visual processing of image statistics via natural scenes, however, is daunting: the statistics of natural scenes constitute a high-dimensional set of parameters that are intertwined in a complex fashion. Recent results concerning high-order statistics of natural scenes (Tkacik et al., PNAS 2010) suggest a way to make this problem tractable. This analysis showed that only certain kinds of high-order statistics are informative about natural scenes; other kinds of statistics can be estimated accurately from simpler quantities and, additionally, are not salient to human observers. Based on this identification of informative statistics, we construct a model “texture space” of binary images, in which textures are specified by the frequency of the colorings of 2 × 2 blocks of pixels. Once these local statistics are specified, long-range statistics are chosen to make the textures as random as possible. The resulting texture space has 10 dimensions. Its coordinates consist of first-, second-, third- and fourth-order image statistics, each of which is visually distinctive. We used a standard segmentation task (4-alternative forced-choice) to determine the perceptual salience of the individual statistics and how they interact. Results are strikingly consistent across N = 6 observers, both qualitatively and quantitatively. With regard to individual statistics, thresholds for the first-, second-, third-, and fourth-order statistics are in a ratio of approximately 1:2:5:4. With regard to interactions, we find (i) that statistics interact across orders, and (ii) that third-order statistics are pooled across orientations to a greater extent than second-order statistics. Together, these studies map out the geometry of a perceptual space, and provide constraints for models of the neural computations that generate it.
This PDF is available to Subscribers Only