A sparse representation based approach for super-resolution reconstruction over geometric dictionaries and consistency of gradients

Recent years has witnessed an increasing interest in handling the issue of single image super-resolution (SISR) reconstruction. Many researches have demonstrated that the sparse representation based approaches, which rely on the ideal that image patches are assumed to have brief representations when expressed in the proper learned dictionaries, can lead to the state-of-the-art performance. The SISR quality via these algorithms depends strongly on whether the utilized dictionaries can describe the potential high resolution image well, therefore dictionaries learning (DL) is of the greatest significance among the procedures. In this paper we determine to utilize a kind of geometric structure based image patches clustering method combined with K-SVD (a DL algorithm) to obtain a better set of geometric dictionaries. To make further progress, we introduce one regularization term, consistency of gradients, into the framework. This term can better preserve the edges in the reconstructed image, making them sharper. Extensive experiments on natural images indicate that our proposed method outperforms some state-of-the-art counterparts in terms of both numerical indicators and visual quality.