Super-Resolution of Video Using Deformable Patches

We introduce the external examples to address the video super-resolution problem. Instead of using sub-pixel complementary information or self-similar examples, we propose the concept that the high frequency video details could be estimated from the external examples effectively. We prepare the high resolution dictionary by randomly selecting patches from the external high resolution images. For each low resolution patch in current frame, we select its neighboring similar patches within current and adjacent frames as the adaptive neighborhood. Their corresponding HR patches in the dictionary are chosen as the deformation candidates. Then we deform them to fit the LR patch. These patches are weighted combined and the final high resolution image is reconstructed by imposing the global constraints. Experiments show that our method outperforms the state-of-art methods.