Video super-resolution using multiple complementary priors

Video super-resolution (SR) is an inverse problem, which has gained much attention in these years. One of the core issues is how to better suppress noise and better preserve the edge. Multi-non-local regularization (MNLR) algorithm is efficient to reduce noise by utilizing the useful information from the correlated frames, but it might also cause some loss of high frequency at the same time. Multi-scale expected patch log likelihood (MSEPLL) and steering kernel regression (SKR) are both fine priors to describe the image characteristics, and could be used to recover details for low resolution images. However, their ability to suppress noise is slightly weaker than MNLR, which means they might preserve some details that MNLR method couldn't recover. Therefore, we consider using a residual framework to extract the useful information from the three reconstructed results. Simply, the images reconstructed by SKR and MSEPLL respectively minus the images reconstructed by MNLR method, and get two different residuals, then use different filters to process them to extract useful information. Finally, adding those processed residuals to the images reconstructed by MNLR method in an adaptive weighting way. The effectiveness of our proposed approach is demonstrated with experiments on video sequences.