Bayesian Comparison of Models for Images

Alex H Barnett, St. John's College

J C MacKay, Cavendish Laboratory

Abstract

Probabilistic models for images are analysed quantitively using Bayesian hypothesis comparison on a set of test image data sets. The aim is to produce models which can be used as better priors for image reconstruction problems.

The types of model vary from the simplest, where spatial correlations in the image are irrelevant, to more complicated ones based on a radial power law for the standard deviations of the coefficients produced from Fourier and Wavelet Transforms. This report's results imply that the Fourier model is the most successful, as its evidence is conclusively the highest. It is shown how the radial power law ties in with the statistically self-similar fractal nature of many images. Discussion of the invariances of the models, and theoretical analysis of their relative performance leads to suggestions and ideas for further investigations.


MaxEnt 94 Abstracts / mas@mrao.cam.ac.uk