Marshall F. Tappen, University of Central Florida
Image enhancement algorithms have seen tremendous improvement as they have adopted non-Gaussian models of natural images. Models like the Field of Experts model or Gaussian Scale Mixture model are designed to model the heavy-tailed distributions commonly seen when natural images are filtered. This talk will focus on the counter-intuitive results showing that a Gaussian model can actually outperform heavy-tailed distributions on image enhancement tasks. The benefit of choosing a Gaussian model is that learning and inference are both greatly simplified due to convenient properties of the Gaussian.
