Progressive low/high-resolution Space Attention Fusion Network for Single Image Super-Resolution

Abstract

The general single image super-resolution methods mainly extract features from the high-resolution (HR) space by the pre-upscaling step at the beginning of the network or from the low-resolution (LR) space before the post-upscaling step at the end of the network. The former way requires high computation as well as misleading the network by wrong artificial priors. The latter way cannot learn mapping well by only conducting simple operations in HR space. In this paper, we aim to utilize the features from LR and HR space more efficiently and propose the novel network, which applies a frequency-slicing mechanism to divide features into LR and HR space, a direction-aware fusion residual group to extract distinctive features in LR space and an attention fusion module to recalibrate features in HR space. The experimental results demonstrate that our model is superior to the state-of-the-art methods upon quantitative metrics and visual quality.