Edge-directed single image super-resolution via cross-resolution sharpening function learning

Abstract

Edge-directed single image super-resolution methods have been paid more attentions due to their sharp edge preserving in the recovered high-resolution image. Their core is the high-resolution gradient estimation. In this paper, we propose a novel cross-resolution gradient sharpening function learning to obtain the high-resolution gradient. The main idea of cross-resolution learning is to learn a sharpening function from low-resolution, and use it in high-resolution. Specifically, a blurred low-resolution image is first constructed by performing bicubic down-sampling and up-sampling operations sequentially. The gradient sharpening function considered as a linear transform is learned from blurred low-resolution gradient to the input low-resolution image gradient. After that, the high-resolution gradient is estimated by applying the learned gradient sharpening function to the initial blurred gradient obtained from the bicubic up-sampled of the low-resolution image. Finally, edge-directed single image super-resolution reconstruction is performed to obtain the sharpened high-resolution image. Extensive experiments demonstrate the effectiveness of our method in comparison with the state-of-the-art approaches.