Lightweight CNN-based Models for Masked Face Recognition

Covid-19 has become one of the most threatening diseases in the world. Therefore, governments and health organizations asked people to wear masks in order to control the spread of such threatening diseases. During this epidemic, people must obey the rules of wearing masks that play a vital role in controlling the outbreak of COVID-19. In order to identify the persons who do not wear the mask, an automated recognition system is discussed in this paper. This model automatically detects masked faces using deep learning techniques; TensorFlow, and Keras. The proposed method effectively differentiates between masked faces and unmasked faces to help governments, companies, and organizations monitoring and detecting who broke the rule of wearing masks. The implementation of this model is based on three pre-trained models; MobileNetV2, DenseNet, and NASNetMobile. In terms of accuracy, MobilenetV2 reached 0.9859, while DenseNets and NASNetMobile reached 0.9852 and 0.9758 respectively. The contribution of this article is to present the feasibility of using a lightweight model (i.e. NasNetMobile) that can be implemented even on low-resource devices (i.e. mobile) to recognize masked faces efficiently.