A blind face inpainting method based on image quality prior

Abstract

A blind face restoration method based on image quality priors addresses the issue of insufficient restoration quality in the field of portrait restoration. By introducing image quality prior information, the restoration algorithm is endowed with image quality perception capabilities, thereby significantly improving the image quality of the restored result during the inference stage. A dual-codebook architecture is designed, retaining diverse and high-quality facial features respectively. Combined with a codeword generation module based on quality prior conditions, and optimizing the Transformer network, high-quality feature codeword sequences are generated based on the quality score of the input image during the codebook lookup stage. Furthermore, the model output space is limited by discrete codebook priors to reduce the generation of adversarial examples, and the model parameters are directly optimized by maximizing the No-Reference Image Quality Assessment (NR-IQA) score, further improving the visual perception quality of the restored image. The restoration results of this invention outperform existing technologies in terms of sharpness, facial detail, and color naturalness, and can be applied to various restoration algorithms without introducing additional inference overhead.

Images