Face Restoration Network with Feature Prior

Recent works on blind face restoration mainly focus on reference-based methods, which made great progress in recovering high-frequency details and realistic texture from the real world low-quality (LQ) images. However, the multi-scale trait of LQ images is not fully utilized with these methods. Extra face reference also takes up much resources and brings redundant model parameters. In this paper, we introduce the face restoration network with feature prior (FP-FRN) consisting of an adversarial network with a multi-scale feature extraction network which utilizes the multi-scale facial feature to preserve low-level facial characteristics and predict high-level details. Compared to other state-of-the-art approaches, i.e., DFDNet, PSFR-GAN, out FP-FRN generates more realistic texture details and better preserved the low-level feature of the LQ images such as color and shape. As demonstrated by experiments on datasets of synthesized and real LQ images, FP-FRN is superior over other methods.