We developed a learning-based super-resolution algorithm that can achieve 16-fold improvement in resolution. We use a database of facial expression videos as a domain-specific prior, and pose the problem of resolution enhancement as one of probabilistic inference. Our results show that by exploiting both spatial and temporal consistencies, hallucinated videos can be rendered more stable and accurate.
Preliminary Results [CVPR '04 Paper]
Note: Video demonstrations available at http://www.cs.cmu.edu/~dedeoglu/thesis/