Aspects of Image Restoration Using Gibbs Priors: Boundary Modelling, Treatment of Blurring, and Selection of Hyperparameters

Valen E. Johnson, Chin-Tu Chen, Xiaoping Hu, Wing H. Wong
Duke University, University of Chicago

Nov 30 1989

We propose a Bayesian model for the restoration of images based on counts of emitted photons. The model treats blurring within the context of an incomplete data problem and utilizes a Gibbs prior to model the spatial correlation of neighboring regions. The Gibbs prior includes line sites to account for boundaries between regions, and the line sites are assigned continuous values to permit efficient estimation using a method called iterative conditional averages. Additionally, the effect of blurring in masking differences between images and the effects of missecifying the amount of blurring are discussed.


Bayesian inference, data augmentation, cross-validation, imcomplete data, EM algorithm


PDF icon 1990-18.pdf