Abstract
Natural images are widely used as stimuli in studies of both lower and higher level visual functions. A commonly used method to distinguish effects related to the second order statistics of natural images from those related to higher order ones is to produce phase-randomized versions of the original stimuli, but with a similar power spectrum, and use them as controls. This approach, however, has major shortcomings. Phase randomization may produce a wider distribution than the original one, possibly with values of intensity outside of the range of the display. Rescaling the values of a randomized image to fit into the desired range would inevitably change the power spectrum as well. In particular, for color images, phase randomization is typically applied to each of the color channels while keeping the original phase differences between channels. With this traditional approach, the 3D joint distribution of original colors is not preserved and could result in substantial differences in color appearance of the phase randomized images which limit their applicability as control stimuli in color vision research. We introduce a method for image randomization that is based on an iterative approximation scheme (a generalization of the IAFFT algorithm; Schreiber & Schmitz, 1996) and which closely preserves the power spectrum of the original image and its exact intensity or 3D color distribution. Applying our algorithm on a collection of natural images revealed that imposing distribution constraints on phase randomized color images results in images with notably closer appearance to that of the original image, not only in color but also in their coarse spatial structure. Our observations suggest that the perceptual information content of first and second order statistics of natural images, when considered jointly, is greater than has been previously appreciated and our proposed method provides means for a reevaluation of their utility in biological vision.
Meeting abstract presented at VSS 2012