Coupled directionally structured dictionaries for single image super-resolution

This paper presents a selective sparse coding algorithm over directionally structured dictionaries learned by using a coupled K-singular value-decomposition (K-SVD) algorithm for single image super-resolution. For a given patch, super-resolution is achieved by enforcing the invariance of the sparse representation coefficients across various scales, and by considering that a sparse representation of a low-resolution patch is being equal to that of a high-resolution patch. The coupled K-SVD algorithm is implemented for the training phase which helps to enforce the similarity between sparse coefficients of the high-resolution and low-resolution patches. Dictionary learning of data is structured into three clusters based on correlation between the patches and already developed horizontal, vertical, and one non-directional template. Coupled dictionaries are learned using the coupled K-SVD algorithm. At the reconstruction phase, each low-resolution patch is correlated with a set of templates for the designed clusters, and that cluster is selected which gives the highest correlation. Then, a pair of dictionaries of that cluster is used for its reconstruction. The proposed algorithm is compared with earlier work, including the currently top-ranked superresolution algorithm. By the proposed mechanism the quality of representation is improved by recovering the directional features more accurately.