An adaptive regression based single-image super-resolution

Image super-resolution (SR) is an important topic of low-level computer vision and is widely used in different fields. In this paper, a novel single-image SR method, which integrates image clustering, sparse representation and linear regression is proposed. Existing global regression based methods usually assume that the corresponding coefficients of HR and LR image patches are equal, which cannot be strictly guaranteed in practice and possibly leads to inaccurate coefficient estimation. In order to adapt the regression model to different types of patches, clustering operations are applied on the training patches to calculate the coefficients of the HR and LR training patches in each class under the same pair of dictionaries. Then the projection relationship between the coefficients of HR and LR training patches in each class is obtained by solving a ridge regression problem. From the experimental results, our algorithm demonstrates better results in both qualitative and quantitative aspects and the computational speed of our methods is noticeably less than other competitive methods.