UFSRNet: U-shaped face super-resolution reconstruction network based on wavelet transform

Aiming to address the issues of excessive parameters and inadequate recovery of facial details in certain existing face super-resolution networks, we propose a U-shaped face super resolution reconstruction network based on wavelet transform. Firstly, a novel Refined Feature Extraction Block (RFEB) is proposed in the Down-sampling Unit, which uses two depth-separable convolution blocks as the main branch and introduces a feature calibration path branch and a residual branch to perform refined feature extraction of the original face images. Secondly, in order to further reduce the number of network parameters, a novel Double Branch Distillation Fusion Block (DBDFB) is designed, which uses two branches to process the features extracted in the down-sampling stage respectively. Finally, Discrete Wavelet Transform (DWT) and Inverse Discrete Wavelet Transform (IDWT) are used to extract and retain high-frequency detail information of face images. Quantitative and qualitative experiments show that our method outperforms state-of-the-art face super-resolution algorithms using only a few parameters. The source codes of the proposed method are available at https://github.com/Aichiniuroumian/UFSRNet.