Single Image Super Resolution Via Adaptive Group-based Sparse Domain Selection

Local smoothness and nonlocal self-similarity of natural images are two main priors in single image super resolution (SISR) problem. Although local sparsity is efficiently utilized to describe the local smoothness, but ignoring the correlation between the sparse representation coefficients of similar patches can lead to inaccurate sparse coding coefficients. In this paper, we propose the method that enforce the local smoothness and nonlocal self-similarity by sparse representation in a unified framework, called adaptive group-based sparse domain selection (A-GSDS). Nonlocal patches with similar structures are exploited and stacked in the form of matrix as the basic unit of sparse representation called group. These groups are converted into a column vector, each column selects the best fitted PCA sub dictionary which is learned from the training data. The sparse coding process for each column in the domain of group leads to find sparse vectors which can be easily estimated by the selected orthogonal sub dictionaries. To further improve the performance of the group-based sparse representation, we use nonlocal means regularization term.  Extensive experimental results validate the effectiveness of the proposed method comparing with the state-of- the-art algorithms.