Return of reconstruction-based single image super-resolution: A simple and accurate approach

Learning-based methods have been the mainstream in single image super-resolution (SISR) over the past decade or so. This paper, however, returns to the reconstruction perspective to SISR, and studies the possibility of achieving competitive or even better SISR performance compared with state-of-the-art learning-based approaches in the literature. Specifically, through the (fast) iterative shrinkage-thresholding algorithm, a general, simple, yet accurate SISR framework is proposed by embedding off-the-shelf image denoising algorithms into the reconstruction process rather than struggling to develop advanced image priors. Numerous experiments on the benchmark dataset are performed accompanied by comparisons against a few recent learning-based methods, e.g., A+, SRCNN, etc. The results have demonstrated well the rationality and superiority of the proposed framework in terms of the super-resolution accuracy as image filtering schemes such as BM3D, WNNM, and other possible candidates, are used. Note that, learning-based filtering schemes can be also plugged into our framework, which hence leads to hybrid reconstruction/learning-based SISR approaches, demonstrating the flexibility as well as the potential of the proposed framework.