Image super-resolution based on adaptive cosparse regularisation

A novel regularised image super-resolution algorithm is proposed, building on the emerging cosparse or analysis sparse prior models, which are important complementary alternatives to the widely used synthesis sparse counterpart. Moreover, to achieve adaptivity to the varying local structures of natural images, the patch space is partitioned into meaningful subspaces by clustering and learn analysis sub-dictionary for each cluster are partitioned, which are performed online and iteratively based solely on the current available image information, for maximum generality and flexibility. In addition, non-local feature self-similarity is incorporated for further reconstruction quality enhancement. Experimental results show that the proposed approach gives favourable results with respect to the state-of-the-art methods.