Deep Learning Mask Face Recognition with Annealing Mechanism

Face recognition (FR) has matured with deep learning, but due to the COVID-19 epidemic, people need wear masks outside reduce risk of infection, making FR a challenge. This study uses FaceNet approach combined transfer learning using three different sizes validated CNN architectures: InceptionResNetV2, InceptionV3, and MobileNetV2. With addition cosine annealing (CA) mechanism, optimizer can automatically adjust rate (LR) during model training process improve efficiency in finding best solution global domain. The mask face (MFR) method is accomplished without increasing computational complexity existing methods. Experimentally, models CA mechanism have better performance than fixed LR, step exponential accuracy reach practical level at about 93%.