FSENet: Frequency Separation Enhancement Network for Super-Resolution

Abstract

Developing a lightweight super-resolution network that can achieve high performance has significant value, but remains under-explored. Many existing methods are devoted to generating rich details, thereby improving the visual quality of single image. However, it may introduce errors or redundant information. To overcome these challenges, we propose a lightweight Frequency Separation Enhancement Network (FSENet) that is focus on both high and low frequency information. Furthermore, we introduce a novel Frequency Gradient-based Perceptual Loss (FGPL), specifically designed to enhance the high-frequency components of the super resolution results. The experimental results demonstrate that the proposed FSENet achieves comparable PSNR/SSIM performance with state-of-the-art lightweight SR models, while achieving a processing speed of 107.8 FPS. This performance surpasses that of all previously reported methods.