Fast single image SR via dictionary learning

In this study, the authors propose a fast method for single image super‐resolution (SR). The relation between high‐resolution (HR)/low‐resolution (LR) patches is learned using the input image and a down‐sampled version. They divide the input image into a number of equal blocks. For each image block, a pair of HR/LR dictionaries, using informative patches, are constructed. For each patch in the input image, an HR dictionary is constructed by concatenating the HR dictionary which it belongs and the HR dictionaries of eight neighbouring blocks. In the same manner, an LR dictionary for each patch is constructed. They represent each patch from the input image using a linear combination of atoms in its LR dictionary. Then, using the same combination for atoms in the HR dictionary of patch, the SR version for the patch is constructed. The computational complexity of the proposed method is considerably low because, in contrast to most of the existing methods, no learning phase, for building the dictionaries of a patch, is required. The experimental results of the proposed method is significantly faster than existing methods, whereas the performance in terms of peak signal‐to‐noise ratio and structural similarity criterions is comparable with the existing methods.