Super-resolution of very low-resolution face images with a wavelet integrated, identity preserving, adversarial network
作者：
Highlights：
• The network minimizes a combination of loss functions to make a balance between distortion and perception.
• The method predicts wavelet coefficients to improve reconstruction of high-frequency components.
• The wavelet-enriched features are concatenated with original features as complementary information.
• The GAN framework is utilized to enhance the perceptual quality.
• An identity loss is aggregated to help the SR network preserve facial identity during hallucination.
摘要
•The network minimizes a combination of loss functions to make a balance between distortion and perception.•The method predicts wavelet coefficients to improve reconstruction of high-frequency components.•The wavelet-enriched features are concatenated with original features as complementary information.•The GAN framework is utilized to enhance the perceptual quality.•An identity loss is aggregated to help the SR network preserve facial identity during hallucination.
论文关键词：Super-resolution,Wavelet prediction,Generative Adversarial Networks,Face Hallucination,Identity preserving,Perceptual quality
论文评审过程：Received 28 September 2021, Revised 7 March 2022, Accepted 27 May 2022, Available online 4 June 2022, Version of Record 23 June 2022.
论文官网地址：https://doi.org/10.1016/j.image.2022.116755
