Deep-learning-based face recognition for worker access control management in hazardous areas

Abstract

Face recognition (FR) technology, which combines computer vision and artificial intelligence, has recently attracted signif-icant attention as a means of identification. Among biometric technologies, FR technology is used in various fields because it does not require physical contact and is hygienic and convenient. Generally, FR processes use imaging equipment to extract facial feature data representing human faces. One can recognize faces by matching the extracted data to facial feature data stored in a database. In this study, we compared the performances of existing deep-learning-based face detection algorithms (i.e., dlib and the single-shot multi-box detector Mobilnet V2) and FR algorithms (i.e., visual geometry groups and ResNet), and developed new FR algorithms, which are crucial for worker access control systems in hazardous regions. To analyze field applicability, we attempted to implement FR algo-rithms with high prediction accuracy in various scenarios (e.g., subjects wearing helmets, protective glasses, or both). We applied regularization to improve the performance of the implemented algorithms. Additionally, related data were collected and analyzed to recognize the number of people wearing masks. The results of recognizing the number of people wearing masks were obtained. These results will support future research on safety issues in the manufacturing industry and the use of face and image recognition techniques.