Learning to estimate missing high-resolution details

William T. Freeman

We address the super-resolution problem: how to estimate missing high spatial frequency components of a static image. We seek to estimate missing high-resolution details from a given low-resolution image. We use an example-based approach, building a database of corrsponding high- and low-resolution image information. The database from a generic set of training images applies well to estimate missing details from images of different subjects. Given a new low-resolution image to enhance, we select from the training data a set of candidate high-frequency patches for each low-resolution input patch. We define a Markov random field for which the most probable state is the desired estimated high resolution image. We find an approximation to the most probable state using Bayesian belief propagation, or other approximations. The algorithm maintains sharp edges, and makes visually plausible guesses in regions of texture.