Jonathan D. Victor¹, Mary M. Conte¹; ¹Weill Cornell Medical College

Processing of visual texture plays an important role in separating figure from ground, and recognizing objects and materials. Visual texture is also a convenient domain for probing the structure of complex perceptual spaces, as texture appearance is the result of different types of low-level features, including contrast, orientation, edges, and corners – and textures can be synthesized in which these features can be independently controlled. Previously we used these texture properties to characterize the nonlinear transformation between the perceptual space supporting texture discrimination thresholds and the space used for suprathreshold similarity judgments. The former is well-modeled by a Euclidean space whose coordinate axes correspond to low-level features and the nonlinear transformation warps those axes and alters the angles between them. Nevertheless, this nonlinear transformation had a dimension-preserving property: for a family of textures with k parameters (typically k=4, with parameters including first, second, and higher-order image statistics) suprathreshold perceptual distances could be largely accounted for by Euclidean distances in a k-dimensional space. To test this property stringently, we obtained texture similarity judgments in 6 subjects and 9 sets of texture parameters. For individual subjects and parameter sets, while the similarity judgments were accounted for by a Euclidean distance, the dimension-preserving fit was imperfect. We analyzed deviations from the fit via shuffle tests that assessed whether addition of further dimensions resulted in a statistically significant decrease in the unexplained variance. This analysis revealed that deviations from the dimension-preserving model were highly consistent both across subjects and texture parameter sets, and supported the addition of at least two dimensions to the space. These findings suggest a distinctive combination of characteristics for the representation of suprathreshold texture similarity: parametric families of textures occupy a low-dimensional subset in a higher-dimensional space, and within this space, similarity corresponds to Euclidean distance.

Acknowledgements: NIH EY07977, Fred Plum Fellowship in Systems Neurology and Neuroscience