Super-resolution using neighbourhood regression with local structure prior

Learning-based super-resolution (SR) imaging has been extensively studied. In this study, a novel reconstruction method is proposed for SR that employs the geometrical structure of an image as well as the statistical priors based on clustering and local structure priors. In this approach, the training samples are divided into numerous overlapping sub-patches, and the sub-patches are then grouped into different clusters, where anchored centres for different subspaces are defined by training joint dictionaries. Variable low-resolution to high-resolution mappings are then learned according to the local similarities. Finally, the desired high resolution patch can be reconstructed by using the multiple learned local relationships corresponding to its sub-patches. Experimental results demonstrated the superior performance of the proposed method.