Realistic Face Masks Generation Using Generative Adversarial Networks

Abstract

Understanding facial expressions is important for the interactions among humans as it conveys a lot about the person's identity and emotions. Research in human emotion recognition has become more popular nowadays due to the advances in the machine learning and deep learning techniques. However, the spread of COVID-19, and the need for wearing masks in the public has impacted the current emotion recognition models' performance. Therefore, improving the performance of these models requires datasets with masked faces. In this paper, we propose a model to generate realistic face masks using generative adversarial network models, in particular image inpainting. The MAFA dataset was used to train the generative image inpainting model. In addition, a face detection model was proposed to identify the mask area. The model was evaluated using the MAFA and CelebA datasets, and promising results were obtained.