Sparse representation based single image super-resolution with low-rank constraint and nonlocal self-similarity

Single image super-resolution reconstruction (SISR) plays an important role in many computer vision applications. It aims to estimate a high-resolution image from an input low-resolution image. In existing reconstruction methods, the nonlocal self-similarity based sparse representation methods exhibit good performance. However, for this kind of methods, due to the independent coding process of each image patch to be encoded, the global similarity information among all similar image patches in whole image is lost in reconstruction. As a result, similar image patches may be encoded as totally different code coefficients. Considering that the low-rank constraint is better at capturing the global similarity information, we propose a new sparse representation model, which concerns the low-rank constraint and the nonlocal self-similarity in the sparse representation model simultaneously, to preserve such global similarity information. The linearized alternating direction method with adaptive penalty is introduced to effectively solve the proposed model. Extensive experimental results demonstrate that the proposed model achieves convincing improvement over many state-of-the-art SISR models. Moreover, these good results also demonstrate the effectiveness of the proposed model in preserving the global similarity information.