Single image super-resolution by combining self-learning and example-based learning methods

In this paper we propose a novel method for single image super-resolution (SISR) by combining self-learning method and example-based learning method. The self-learning method we used which is proposed by Zeyde et al. (2012) has the ability to scale-up a single image from the given low-resolution (LR) image itself by learning a dictionary pair directly from the given LR image (as high-resolution image) and its scaled-down version (as LR image). This is the so-called boot-strapping method in Zeyde, Elad, Protter (Lect Notes Comput Sci 6920:711–730, 2012). With the output image obtained by the boot-strapping method and the original high-resolution (HR) image, we can get a super-resolution image by learning the sparse representation model proposed in Na, Jinye, Xuan, Xiaoyi (Multimed Tools Appl 74:1997–2007, 2015). Our combined approach shows to perform better even when there are little training example images. A number of experimental results on true images show that our method gains both visual and PSNR improvements.