Complementing SRCNN by Transformed Self-Exemplars

Super-resolution algorithms are used to improve the qualityand resolution of low-resolution images. These algorithms can be dividedinto two classes of hallucination- and reconstruction-based ones. Theimprovement factors of these algorithms are limited, however, previousresearch [10], [9] has shown that combining super-resolution algorithmsfrom these two different groups can push the improvement factor further.We have shown in this paper that combining super-resolution algorithmsof the same class can also push the improvement factor up. For thispurpose, we have combined two hallucination based algorithms, namelythe one found in Single Image Super-Resolution from Transformed Self-Exemplars [7] and the Super-Resolution Convolutional Neural Networkfrom [4]. The combination of these two, through an alpha-blending, hasresulted in a system that outperforms state-of-the-art super-resolutionalgorithms on public benchmark datasets.