Joint face completion and super-resolution using multi-scale feature relation learning

Previous research on face restoration often focused repairing specific types of low-quality facial images such as low-resolution (LR) or occluded images. However, in the real world, both above-mentioned forms image degradation coexist. Therefore, it is important to design a model that can repair are LR and simultaneously. This paper proposes multi-scale feature graph generative adversarial network (MFG-GAN) carry out contexts which modes coexist, also with single type degradation. Based GAN, MFG-GAN integrates convolution pyramid networks restore non-occluded high-resolution The uses set customized losses ensure high-quality generated. In addition, we designed an end-to-end format. We conduct experiments general expression restoration. Experimental results public-domain databases show proposed approach outperforms state-of-the-art methods performing super resolution (up 4× 8×) completion simultaneously recover better details.