Human visual coding and the third-order statistics of natural scenes

Mitchell Thomson

Many studies have investigated the possibility of a `perceptual match' between the properties of the human visual system and those of natural scenes, but these analyses have so far been largely restricted to second-order statistical measures. Demonstrations of phase randomization and phase quantization imply that it is the organization of the Fourier phase spectrum which is the critical determinant of perceived scene structure. Second-order image-analysis techniques are effectively phase-blind; to quantify phase relationships one must compute higher-order statistics of image data. It can be shown that the triple correlation function determines a bandlimited image of fixed position uniquely and completely, and third-order image measures can predict human performance in psychophysical image-discrimination tasks. The present work extends existing 1-D results to 2-D image data: a third-order coherence measure was computed for a variety of natural images before and after Fourier-phase randomization. A wavelet-based technique was used to assess the effect on this coherence measure of processing the images through independent spatial-frequency channels. The results suggest that although a constant-octave-band representation may encode equal energy in equal octaves, it does not encode equal (third-order) structure in equal octaves: the predicted optimal pattern of spatial-frequency tuning appears instead rather more similar to existing experimental data on cortical-cell bandwidths.
Back to the Natural Scenes Meeting Agenda