Deep Learning-based Face Mask Usage Detection on Low Compute Resource Devices

The advent of the COVID-19 pandemic has brought several never-before-seen changes in the daily lives of people around the globe. As a way to curb the spreading of the disease, wearing face masks has become mandatory in the majority of public places. To solve the necessity of face mask detection in such situations, there have been only a handful of research endeavors up to this date. Computer vision has advanced multi-fold with the advent of AlexNet architecture. With a motivation to go deeper with the neural network architecture, the concept of Depthwise Separable Convolutions and projection layer was developed in MobileNetV1. In this work, a novel lightweight deep learning model based on Single Shot Detector (SSD) MobileNetV2 architecture is proposed for face mask detection using images and video streams of crowds aiming its utilization on low compute resource environment. An open benchmark face mask dataset, with 4095 images including masked and no mask images, is utilized to train the model for detection. The model is initialized using transfer learning with the freezing of base layers. The proposed methodology can efficiently aid in tracking and enforcing social distancing rules in crowded places with the use of surveillance cameras. On the different benchmarks that we have tested, the model proved to be highly successful and has achieved an accuracy rate of 99.39% and an F1 score of 0.995.