A self-learning image super-resolution method via sparse representation and non-local similarity

It is difficult to design an image super-resolution algorithm that can not only preserve image edges and texture structure but also keep lower computational complexity. A new super-resolution model based on sparsity regularization in Bayesian framework is presented. The fidelity term restricts the underlying image to be consistent with the observation image in terms of the image degradation model. The sparsity regularization term constraints the underlying image with a sparse representation in a proper dictionary. The non-local self-similarity is also introduced into the model. In order to make the sparse domain better represent the underlying image, we use high-frequency features extracted from the underlying image patches for sparse representation, which increases the effectiveness of sparse modeling. The proposed method learns dictionary directly from the estimated high-resolution image patches (extracted features), and the dictionary learning and the super-resolution can be fused together naturally into one coherent and iterated process. Such a self-learning method has stronger adaptability to different images and reduces dictionary training time. Experiments demonstrate the effectiveness of the proposed method. Compared with some state-of-the-art methods, the proposed method can better preserve image edges and texture details.