Item

Abstract

Several studies have demonstrated that the power spectra of natural image ensembles scale as f−α. A stronger claim that has been made is that the power spectra of single natural images typically also scale as f−α. Results are presented that challenge this latter claim. The results are based on a method for estimating large-scale structure in single images that compares aliasing artifacts produced by image windows of different shape. Failures of f−α scaling are found at large scales in many natural images. These failures cannot be accounted for by f−α scaling models such as a linear superposition model or a model based on two-dimensional occlusions in the image plane. The results imply that claims about f−α scaling in single natural images have been exaggerated. The results also offer insight into why such failures of f−α scaling occur.