A method for generating training data for a protective face mask detection system

Monitoring and evaluation of the safety level of individuals is one of the most important problems of the modern world, which was forced to change due to the emergence of the COVID-19 virus. To increase the safety level of individuals, new information technologies are needed that can stop the spread of infection by minimizing the threat of outbreaks and monitor compliance with recommended measures. These technologies, in particular, include intelligent tracking systems of the presence of protective face masks. For these systems, this article proposes a new method for generating training data that combines data augmentation techniques, such as Mixup and Insert. The proposed method is tested on two datasets, namely, the MAsked FAce dataset and the Real-World Masked Face Recognition Dataset. For these datasets, values of the unweighted average recalls of 98.51% and 98.50% are obtained. In addition, the effectiveness of the proposed method is tested on images with face mask imitation on people's faces, and an automated technique is proposed for reducing type I and II errors. Using the proposed automated technique, it is possible to reduce the number of type II errors from 174 to 32 for the Real-World Masked Face Recognition Dataset, and from 40 to 14 for images with painted protective face masks.