Deep Residual Network With Enhanced Upscaling Module for Super-Resolution

Jun-Hyuk Kim, Jong-Seok Lee; Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2018, pp. 800-808

Single image super-resolution (SR) have recently shown great performance thanks to the advances in deep learning. In the middle of the deep networks for SR, a part that increases image resolution is required, for which a sub-pixel convolution layer is known as an efficient way. However, we argue that the method has room for improvement, and propose an enhanced upscaling module (EUM), which achieves improvement by utilizing nonlinear operations and skip connections. Employing our proposed EUM, we propose a novel deep residual network for SR, called EUSR. Our proposed EUSR was ranked in the 9th place among 24 teams in terms of SSIM in track 1 of the NTIRE 2018 SR Challenge. In addition, we experimentally show that the EUSR has comparable performance on x2 and x4 SR for four benchmark datasets to the state-of-the-art methods, and outperforms them on a large scaling factor (x8).