ABSTRACT
In this chapter, the authors describe the Bayesian framework for systematically modeling the observed low-resolution images, the unknown high-resolution (HR) image, and the motion and blur parameters. Using this model, they develop two super-resolution (SR)algorithms that jointly estimate the HR image and all algorithmic parameters. The authors demonstrate experimentally that the proposed methods provide HR images with high quality and compare favorably to existing SR methods. They discuss extensions of the model to incorporate the estimation of the motion and blur parameters. The general variational Bayesian analysis can be directly utilized for Gaussian observation models and Gaussian image priors. The errors in estimating the blur and registration parameters cause significant drawbacks in super-resolution, leading to instabilities in the recovery of the HR image and significantly affecting the robustness of the restoration procedures. Some methods utilize robust image estimation methods to alleviate the problems caused by the errors in the motion estimates.
