COVID-19 Face Mask Restoration Using Pix2pix model

ABSTRACT

Accurate facial recognition can effectively help the population combat the disease by offering risk-free phone usage, access controls, etc. In the era of COVID-19, a mask has become a necessity. However, masks may reduce the accuracy of face recognition to some degree. Thus, it is necessary to use deep learning to increase face recognition accuracy by recovering the face with a mask. For this purpose, this study proposed an AI-based model based on Pix2pix and U-net generator for restoring face mask images using the paired image database. In the training step, we used two adversarial models, including one generator and one discriminator. Then they are extended to a conditional model, which will be piped to the Pix2pix algorithm once again. U-Net was built in the training of the generator. The loss curves of generator and discriminators show that as iteration time increases, the loss of fake discriminator becomes lower stably. In contrast, the loss of real discriminator has the same tendency. In the meantime, the loss of generator shows an increased tendency. The result indicates that our model can help build reliable face mask restoration for daily use, which helps to improve the recognition accuracy of the face with a mask.