Role of second- and third-order statistics in the discriminability of natural images

Several studies of natural-image statistics have concluded that the majority of structural image information is represented in the Fourier domain not by the power spectrum, which can be modeled with a power function of spatial frequency, but by the phase spectrum. Psychophysical experiments are described in which human observers were required to discriminate between visual stimuli differing only in their global second-order statistics. The patterns of sensitivity recorded were found to depend on whether these stimuli were natural images or images with random-phase spectra. These results suggest that discrimination performance is related to changes in higher-order image structure; a bandpass model observer based on third-moment statistics was applied to all stimulus images and was found to provide a good fit to the psychophysical data.