The detection of structure in glass patterns: psychophysics and computational models.

Experiments are reported which examine the judgement of the mean orientation of textures composed either of short lines or dipoles (Glass patterns). The effects of element length, density, and orientation variation are described. Psychophysical data are compared with predictions from four schemes for extracting features from Glass patterns: token matching, isotropic filtering, oriented filtering, and "adaptive" filtering (selection of local peak output from multiply oriented filters). Glass patterns are spatially broadband but only contain orientation structure at a narrow range of scales making them suitable for examining how filter size is selected for texture processing. A criterion for scale selection is proposed: that local variation of feature orientation should be minimized. Simulations indicate that neither models using isotropic filtering nor token matching achieve human levels of performance on certain tasks. Adaptive filtering, operating at a scale selected using the criterion described, provides good agreement with the psychophysical data reported and is a practical scheme for deriving features using oriented filters.