An improved joint dictionary training method for single image super resolution

Research on image statistics suggests that image patches can be well represented as a sparse linear combination of elements from an appropriately over-complete dictionary. In this paper, an improved joint dictionary training scheme is introduced for the single image super resolution. By using different weight factors, the scheme balances two dictionaries in the high- and low- resolution spaces in the training to achieve good reconstructed images. A K-SVD algorithm is applied to learn the dictionaries. Sparse representations of low-resolution image patches are used to reconstruct the high-resolution image patches. From the experiment results, the proposed scheme outperforms the classic bicubic interpolation and neighbor embedding learning based method both qualitatively and quantitatively.