基于 Swin Transformer 和 Style-based Generator 的盲人脸修复

盲人脸修复任务是从低质量的图像（例如模糊、噪声和压缩图像）中恢复高质量的图像.由于事先不知道低质量图像的退化类型和退化参数，因此盲人脸修复是一个高度病态的问题，在修复过程中严重依赖各种先验指导.然而，由于面部成分和面部标志等面部先验通常是从低质量图像中提取或估计的，可能存在不准确的情况，这直接影响最终的修复性能，因此难以有效利用这些先验知识.此外，目前的主流方法基本都是依赖ConvNets进行特征提取，没有很好地考虑长距离特征，导致最终结果缺乏连续一致性.本文提出了一种改进的StyleGAN模型，命名为SwinStyleGAN，应用在高级视觉任务上表现出色的Swin Transformer来提取长距离特征，并通过改进后的类StyleGAN合成网络逐步生成图像.本文设计了一个空间注意力转换模块SAT来重新分配每个阶段特征的像素权重，以进一步约束生成器.大量实验表明，本文提出的方法具有更好的盲人脸修复性能.

Blind face restoration based on Swin Transformer and Style-Based Generator

Blind face restoration is the process of restoring a high-quality image from a low-quality image (e.g., blurred, noisy, or compressed image). Since the degradation type and degradation parameters of the low-quality image are unknown, blind face restoration is a highly ill-posed problem that heavily relies on various facial prior such as facial components and facial landmarks during the restoration process. However, these facial priors are typically extracted or estimated from low-quality images, which may be inaccurate, directly affecting the final restoration performance. The current mainstream methods mostly use ConNets for feature extraction and do not consider long-distance features, resulting in a lack of continuous consistency in the final results.The authors propose an improved StyleGAN model named SwinStyleGAN, which uses Swin Transformer to extract long-distance features and gradually generates images through an improved StyleGAN synthesis network.Addtionally, the authors design a Spatial Attention Transformation (SAT) module to reassign pixel weights of each stage feature to further constrain the generator. Experiments show that the proposed SwinStyleGAN in this paper has better blind face restoration performance.