Abstract
Human observers can effortlessly tell the difference between natural images and randomly generated noise. Some elementary statistical features that distinguish natural images can be captured in “naturalistic texture” images generated by a method devised by Portilla & Simoncelli (2000). These textures are easily distinguishable from randomly generated images with identical spectra, but it is unclear what features drive this perceptual sensitivity. We wondered whether certain frequency bands are particularly important for naturalness perception.
We created families of images that span the range from fully naturalistic textures to spectrally matched “noise.” Suitably combined, these create images with naturalistic structure limited to one of three spatial frequency bands: a low-pass spatial frequency band, a high-pass spatial frequency band, or the full spatial frequency spectrum. This allowed us to test how well an observer could discriminate the presence or absence of naturalistic structure, using a three-alternative forced-choice oddity task. As a control, we also had the observer perform a contrast discrimination task using the same base images, but now discriminating the presence or absence of contrast steps within the same low, high, and all-pass spatial frequency bands.
Contrast discrimination was more strongly impaired by the removal of low-frequency contrast steps than of high-frequency steps. Conversely, performance in the texture discrimination task was largely robust to the removal of low-frequency naturalistic structure. These results suggest that our perception of image naturalness may rely on higher spatial frequency bands than those for which contrast sensitivity is highest. The prominent role of high-frequencies in texture discrimination may help explain a recently described deficit in texture sensitivity observed in amblyopic patients (Lee et al. 2017, SfN Meeting Abstract), who typically lack sensitivity to high spatial frequencies.