Blind Super-resolution for Single Image Reconstruction

Image super-resolution reconstructions (SR) require image degradation model (DM) as the prior, however, the actual DM is often unknown in practical applications. In this work, a novel framework is proposed for single image SR, where the explicit DM is unknown. Based on Bayesian MAP, an iteration scheme is adopted to update the reconstructed SR image and the DM estimate. During reconstruction, MRF-Gibbs image prior is incorporated for regularization and example-based machine learning technique is employed to draw the DM estimations back to the potential DM space. The SR resulted images by the proposed method are superior to the ones produced by bicubic interpolation and conventional SR algorithm with incorrect DM, in both aesthetical and quantitative aspects.