A Correct Face Mask Usage Detection Framework by AIoT

ABSTRACT

The COVID-19 pandemic, which affected over 400 million people worldwide and caused nearly 6 million deaths, has become a nightmare. Along with vaccination, self-testing, and physical distancing, wearing a well-fitted mask can help protect people by reducing the chance of spreading the virus. Unfortunately, researchers indicate that most people do not wear masks correctly, with their nose, mouth, or chin uncovered. This issue makes masks a useless tool against the virus. Recent studies have attempted to use deep learning technology to recognize wrong mask usage behavior. However, current solutions either tackle the mask/non-mask classification problem or require heavy computational resources that are infeasible for a computational-limited system. We focus on constructing a deep learning model that achieves high-performance results with low processing time to fill the gap in recent research. As a result, we propose a framework to identify mask behaviors in real-time benchmarked on a low-cost, credit-card-sized embedded system, Raspberry Pi 4. By leveraging transfer learning, with only 4-6 h of the training session on approximately 5,000 images, we achieve a model with accuracy ranging from 98 to 99% accuracy with the minimum of 0.1 s needed to process an image frame. Our proposed framework enables organizations and schools to implement cost-effective correct face mask usage detection on constrained devices.