Regression-based single image super-resolution via adaptive patches

Single image super-resolution (SR) generates a high-resolution (HR) image by estimating the mapping function between image patches of different resolutions. By leveraging the notion of regression, the mapping function estimation task is often transformed into predicting mapping function's derivatives. Although higher-orders of derivative lead to a more accurate mapping function, current algorithms only achieve the first-order derivative estimation, due to the ill-conditioned nature of such estimation problem. By observing that the size of patches not only influences the illness of this estimation problem, but also affects the detail reconstruction in the final HR image, we incorporate an adaptive patch size scheme into single image SR in this paper, so as to facilitate the SR algorithm to detail preservation. Experiments on standard images demonstrate the effectiveness of the proposed method both quantitatively and qualitatively, when comparing to other advanced SR algorithms.