Research on High-Resolution Face Image Inpainting Method Based on StyleGAN

• DOI: 10.3390/electronics11101620
• Publication Date: 2022-05-19T16:25:57Z
• Authors: Libo He; Zhenping Qiang; Xiaofeng Shao; Hong Lin; Meijiao Wang; Fei Dai
• Keywords: Inpainting; Face hallucination

In face image recognition and other related applications, incomplete facial imagery due to obscuring factors during acquisition represents an issue that requires solving. Aimed at tackling this issue, the research surrounding completion has become important topic in field of processing. Face methods require capability capturing semantics expression. A deep learning network been widely shown bear ability. However, for high-resolution completion, training inpainting is difficult converge, thus rendering a problem. Based on study model generation, paper proposes method. First, our method extracts latent vector be repaired through ResNet, then inputs pre-trained StyleGAN generate image. Next, it calculates loss between known part corresponding generated imagery. Afterward, cut new iteratively until number iterations reached. Finally, Poisson fusion employed process last order eliminate difference boundary color information Through comparison analysis two classical recent years CelebA-HQ data set, we discovered can achieve better results 256*256 resolution completion. For 1024*1024 restoration, have also conducted large experiments, which prove effectiveness Our obtain variety repair by editing vector. addition, successfully applied editing, watermark clearing applications without different masks these applications.