Blind Face Restoration Survey

Abstract

The importance of researching methods for blind face restoration (BFR) arises from their potential practical applications in various domains. Examples of such areas include digital art and computer graphics for character face reconstruction and animation, as well as social networks and mobile applications, where they contribute to improving the quality of images and videos. In this paper, we conduct a review of contemporary methods and approaches used for solving the BFR problem. We examine various types of models based on generative adversarial networks, autoencoders, and diffusion models, which have demonstrated significant progress in this field. Specifically, we analyze key aspects such as network architecture, loss functions, quality metrics, and datasets. Furthermore, we discuss the issues and limitations of existing methods, as well as possible directions for future research. In particular, we emphasize the need for developing algorithms that are robust to various degradations and capable of adapting to different lighting conditions, poses, and facial expressions. In conclusion, we provide a systematic comparison of existing methods and summarize their merits and drawbacks.

What the survey covers

The paper is a literature review in Russian, with an English abstract. It runs no new experiments and reports no benchmark numbers of its own.

• Problem statement. Separate degradation models for denoising, deblurring, super-resolution and JPEG artifact removal, then blind restoration as an unknown mix of them.
• Quality metrics. PSNR, SSIM, LPIPS, FID and NIQE, with what each one measures.
• Taxonomy by prior. Methods with no prior, with a geometric prior (landmarks, parsing maps), with a reference prior (component dictionaries, learned codebooks) and with a generative prior (pretrained GANs and diffusion models).
• Method table. 17 methods with prior, architecture and key idea: STUNet, HiFaceGAN, DFDNet, PULSE, SPARNetHD, GFP-GAN, GPEN, PSFRGAN, GLEAN, FaceFormer, RestoreFormer, CodeFormer, VQFR, DDRM, DDNM, DifFace and DR2.
• Datasets. CelebA and CelebA-Test, FFHQ, CASIA-WebFace, VGGFace2, IMDB-WIKI, Helen, WIDER Face, LFW, BioID, AFLW, and the two sets built for BFR: EDFace-Celeb-1M (BFR128) and EDFace-Celeb-150K (BFR512).

The open problems the authors name are restoration under heavy degradation and better recovery of facial texture and detail.