Purchase this article with an account.
Michael Rudd; Edge integration and image segmentation in lightness and color. Journal of Vision 2017;17(10):650. doi: 10.1167/17.10.650.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Previous work has demonstrated that lightness in disk-annulus displays can be explained by a computational model based on the principle of edge integration (Rudd, Front. Hum. Neurosci., 2014). The disk lightness depends not just on the contrast, or luminance ratio, of the disk with respect to its immediate surround, but instead on a weighted sum of luminance steps (in log units) computed along a path from the background, through annulus, to the disk. The idea of summing steps in log luminance to compute lightness makes ecological sense as a means of relating the disk luminance to the background luminance for the purpose of constructing a global reflectance map, starting from information about local changes in luminance that might be encoded by early visual neurons (Rudd, 2014; J. Electron. Imaging, submitted). Edge integration theory explains disk-annulus lightness matching results with great precision, but can the theory generalize to explain surface lightness and color appearance in other contexts? Here I shown how it can account for assimilation in the classic Helson display, and in a new display--the Nebraska election map--in which the yellow color of hatched diagonal lines draw within an otherwise red region bleed into the red. I also present displays that I have constructed by modifying White's display. In these displays, edge integration operates within regions that are first visually segmented according to principles of perceptual organization. The phenomenology of my modified White's displays supports the conclusion that steps in log luminance sum to compute target lightness only within the segmented region to which the target perceptually belongs, suggesting a role of perceptual grouping or scission in depth in White's effect. My new results cannot be explained either by low-level normalized filter models, such as ODOG, or by an edge integration theory that ignores the role of image segmentation.
Meeting abstract presented at VSS 2017
This PDF is available to Subscribers Only