Multispectral Face Mask Compliance Classification During a Pandemic

Abstract

AbstractIn this paper, we investigate the problem of face mask compliance classification in response to the Corona Virus Disease (COVID-19) pandemic. Since the start of the pandemic, many governments and businesses have been continually updating policies to help slow the spread of the virus, including requiring face masks to use many public and private services. In response to these policies, many researchers have developed new face detection and recognition techniques for masked faces. Many of the developed approaches proposed to address the problem of masked face detection have been relatively successful. However, these approaches almost exclusively focus on detecting the presence or absence of someone wearing a face mask. It is understood and broadly discussed in various reports and the media that there are people not always following the suggested guidelines provided by public health authorities for wearing face masks, which include ensuring the face mask properly covers the nose and mouth. To date, very few research publications exist that investigate the capability of modern classification algorithms to efficiently distinguish between masked faces in the visible band that are either compliant or non-compliant with the suggested guidelines. Furthermore, to the best of our knowledge, there is no publication in the open literature that focuses on the investigation of face mask classification in the thermal band. As thermal sensors continue to improve in quality and decrease in cost, surveillance applications using thermal imagery are expected to continue to grow and benefit those organizations considering the automation of face mask detection and compliance. In this study, we propose an investigation on face mask compliance in both the thermal and visible bands. It is composed of the following salient steps; (1) the creation of a multi-spectral masked face database from subjects wearing or not wearing face masks, (2) the augmentation of the generated database with synthetic face masks to simulate two different levels of non-compliant wearing of face masks, and (3) the assessment of a variety of CNN architectures on the previous augmented database to investigate any differences between classifying thermal and visible masked faces. Experimental results show that face mask compliance classification in both studied bands yield a classification accuracy that reaches 100% for most models studied, when experimenting on frontal face images captured at short distances with adequate illumination.KeywordsFace masksCOVID-19PandemicFace mask classificationDeep learningMask complianceThermal imagingMulti-spectral Imaging