Purchase this article with an account.
Mary Conte, Syed Rizvi, Jonathan Victor; Approximately uniform isodiscrimination contours within a perceptual space of local image statistics. Journal of Vision 2015;15(12):774. doi: 10.1167/15.12.774.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Analysis of image statistics is crucial to segmenting a visual scene and characterizing its components. Disentangling the computations used for this analysis is challenging, as image statistics form a high-dimensional domain in which edges, corners, and other local features of natural scenes interact in a complex fashion. We therefore developed a reduced space of synthetic images in which these features can be varied independently. This space consists of binary (black-and-white) images, parameterized by the configurations within 2x2 neighborhoods. Its 10 dimensions capture many of the image statistics that are informative in natural images (Tkačik et al., 2010). Previously (Victor et al., VSS 2013) we showed that at the origin of the space (i.e., for discrimination from randomness), threshold judgments implied a simple combination rule for image statistics: they combined in a quadratic fashion, generating ellipsoidal isodiscrimination contours. Here, we extend this analysis to the periphery of the space and show that the result generalizes. To measure discrimination thresholds, we used a segmentation paradigm. Stimuli consisted of 64x64 arrays of 14-min checks, containing a 16x64 target in one of four locations. Target and background were each defined by structured binary textures, with one chosen at a reference point in the periphery of the space and the other at a parametrically-varied distance from the reference. Thresholds were determined in N=4 subjects from Weibull fits to their psychometric functions. We found that the sizes and shapes of the isodiscrimination contours at peripheral points in the space were similar to the isodiscrimination contours determined, in parallel, at the origin. Thus, over the range tested, perceptual thresholds are determined primarily by the vector difference between image statistics. This simple and approximately Euclidean representation exists in parallel with a highly curved representation (Rizvi et al., VSS 2014) required to account for suprathreshold border saliences.
Meeting abstract presented at VSS 2015
This PDF is available to Subscribers Only