A Quantitative and Qualitative Analysis on a GAN-Based Face Mask Removal on Masked Images and Videos

In 2020 and beyond, there are more and more opportunities to communicate with others while wearing a face mask. Since masks hide the mouth and facial muscles, it becomes more challenging to convey facial expressions to others while wearing a face mask. In this study, we propose using generative adversarial networks (GAN) to complement facial regions hidden by masks on images and videos. We defined the custom loss function that focuses on the error of the feature point coordinates of the face and the pixels in the masked region. As a result, we were able to generate higher-quality images than existing methods. Even when the input was video-based, our approach generated high-quality videos with fewer jittering and pixel errors than existing methods.