Efficient Single Image Super-Resolution via Hybrid Residual Feature Learning with Compact Back-Projection Network

Deep learning methods have achieved state-of-the-art accuracy in single image super-resolution (SISR). Yet, how to achieve good balance between efficiency and accuracy in SISR is still an open issue. While most existing methods learn residual features only in low resolution (LR) space in order for higher efficiency, recent studies show that jointly learning residual features in LR and high resolution (HR) space is more preferred for accurate SISR. In this paper, we propose an efficient SISR method via learning hybrid residual features, based on which the residual HR image can be reconstructed. To fulfill hybrid residual feature learning, we propose a compact back-projection network that can simultaneously generate features in both LR and HR space by cascading up- and down- sampling layers with small-sized filters. Extensive experiments on four benchmark databases demonstrate that our proposed method can achieve high efficiency (i.e., small number of parameters and operations) while preserving state-of-the-art SR accuracy.