A class of multiresolution stochastic models generating self-affine images

Fractal image compression is based on the rather poorly motivated assumptio n that "natural" images exhibit significant affine selfsimilarity. The accu racy of this assumption is evaluated by a comparison between the statistics of natural images and those of a multiresolution stochastic model designed to generate images exhibiting affine self-similarity as assumed by fractal coding techniques. These comparisons suggest that self-affinity does not r epresent a particularly accurate characterization of image statistics.