Mask Removal Algorithm Using GAN Based Models

Recent deep learning network models have been enhanced dramatically to generating images, removing objects, and improving resolution of images. From a pragmatic purpose, we must utilize it to solve some occurrent real-world problems, for example, facial mask removal. People's frequency of wearing masks has increased rapidly due to the pandemic. The objective of this research is to remove the facial mask on face images using generative adversarial network, or GAN, model. We set up three different training process to explore how to remove a facial mask effectively. By comparing among these training processes, we find out that adding an additional binary mask segmentation is conducive to both the generation quality of face region other than mask and inpainting quality of missing mask region. Moreover, our customized GAN model outperforms the Pix2Pix GAN model in both region inpainting.