The role of high-order phase correlations in texture processing

AbstractIsodipole textures are pairs of texture ensembles whose autocorrelations, and hence power spectra, are equal. Examples of readily discriminable isodipole textures are well known. Such discriminations appear to require feature extraction, since the isodipole condition eliminates ensemble differences in spatial frequency content. We studied the effects of phase decorrelation on VEP indices of discrimination of isodipole texture pairs. Phase decorrelation, which ranged from 0.125π radians (slight randomization) to π radians (complete randomization), was introduced in two ways: by independent jittering of each spatial Fourier component, and by a product method, which preserved correlations among certain quadruples of spatial Fourier components, despite pairwise decorrelation. For the even/random isodipole texture pair, independent phase decorrelation greater than 0.5π radians markedly reduced VEP indices of texture discrimination for all check sizes, and eliminated them entirely for check sizes of 8 min or greater. However, the product method preserved texture discrimination signals even with complete pairwise randomization of spatial phases. For the triangle/random isodipole texture pair, both kinds of phase decorrelation eliminated VEP indices of texture discrimination. These results imply that isodipole texture discrimination is based on fundamentally local processing, and not on global Fourier amplitudes—since the phase manipulations which eliminate texture discrimination preserve the Fourier amplitudes. The dependence of the antisymmetric response component (the odd harmonics) on phase decorrelation and texture type is consistent with a previously proposed model for feature extraction, and leads to constraints on how texture processing is modulated by contrast. The limited contribution of global spectral characteristics for small checks is consistent with a previously identified breakdown in scale-invariant processing