Efficient Deep Learning Model for Real-Time Mask Detection on Mobile Devices
By: Zhenyu Lin, Benediction Bora, and Philip Liang
Department: Computer Science
Faculty Advisor: Dr. Zhuwei Qin
Nowadays, the global spread of COVID-19 has changed everyone’s life. With the reopening of countries from COVID-19 lockdown, one of the effective ways to control the spreading of the virus is to wear a face mask. Most governments and institutions have set rules to force wearing a facemask in public and workplaces. However, manual real-time monitoring of face mask wearing for a large group of people is becoming a difficult task as it will cause additional human resources and financial costs. Recently, there are some works has been proposed to automated detect face mask by using deep learning models. However, the deep learning model requires a lot of computational and storage costs. Thus, these models are usually developed on high-performance desktop computers which limits their wide usage in daily life. In this work, we proposed an efficient deep learning compression method that can remove the redundant neurons in the deep learning model to achieve low inference time and high accuracy. The MobileNetV2 is used as a deep learning model for face mask detection and a real-time user interface is built to display the number of people without wearing the mask. The system is implemented on Nvidia Jetson Nano mobile devices, which can be further implemented in real-world surveillance cameras in public areas to check if people are following rules and wearing marks.