Sertis face mask detection pipeline for mask-wearing monitoring in the era of the COVID-19 pandemic: A modular approach
Since the global outbreak of the SARS-Cov-2 (COVID-19) virus in the later part of 2019, more than five million deaths have been reported globally, with many more hospitalized and under critical condition. The SARS-Cov-2 is a respiratory virus and spreads by a mixture of large droplets and airborne particles. Thus, many countries introduced non-pharmaceutical interventions such as mask-wearing, hand-washing, and social distancing to curb its transmission.
Mask-wearing in public areas, particularly, was deemed a stringent measure that spurred controversies leading to its varying acceptance levels among many countries initially. As we gradually settle into the new normal, this measure has become more accepted as an effective instrument in decreasing the spread of COVID-19 disease. Most of the studies evaluating the rate of mask-wearing use the presence of mask mandates or online questionnaires or use one-off documentation of mask-wearing data. These studies, however, have their limitations in terms of accuracy and, more importantly, lack monitoring of the changes in the mask-wearing rate in public. Thus, making monitoring the mask-wearing rate within a population a practical tool for determining public health strategies against the virus.
With the current advancement in machine learning (ML) and computer vision, researchers have developed new techniques that have proved helpful in analyzing and detecting faces with and without masks, as validated by many published reports. However, in reality ML models for detecting faces with masks have not been deployed at a large scale to measure the mask-wearing rate in public. Thus, serving as the primary motivation for the Sertis AI research team to develop a face mask detection ML pipeline for public mask-wearing monitoring.
In this peer-reviewed research, the team constructed a two-step face mask detection pipeline consisting of two main modules: 1) face detection and alignment and 2) face mask classification. The modularity of this pipeline allows us to use existing efficient face detectors such as PyramidKey [1] and RetinaFace [2] and easily update or change either of the face detection or the face mask classification modules. For training the face mask classification model, the team investigated different training techniques on the face mask classification model, i.e., label smoothing [3], aligning with eyes key points, and ignoring the upper half face. They showed this pipeline outperforming several state-of-the-art approaches on benchmark datasets, i.e., AIZOO [4] and Moxa 3K [5] datasets which contain images from various scenarios from close-ups to crowded scenes.
The development of this research was a collaboration between the Sertis Co., Ltd. and the Faculty of Medicine, Ramathibodi Hospital, Mahidol University. From the Sertis AI research team, the project was led by our AI researcher Benjaphan Sommana and included other researchers, namely Samuel W. F. Earp, Ukrit Watchareeruetai, and Ankush Ganguly. Additionally, medical practitioners from the Faculty of Medicine, Ramathibodi Hospital, namely Taya Kitiyakara, Suparee Boonmanunt, and Ratchainant Thammasudjarit, provided active supervision during the development of this research which was successfully accepted and presented at the 19th International Joint Conference on Computer Science and Software Engineering (JCSSE 2022).
This research successfully demonstrates that our proposed pipeline is superior than the state-of-the-art approaches for face-mask detection. This success has marked itself as a milestone in Sertis’ journey to develop state-of-the-art technology not only to be a part of the AI industrial revolution but also to aid in the improvement of healthcare strategies for the betterment of lives.
Read the full research at : https://arxiv.org/abs/2112.15031
Written by: Sertis Vision Lab
Originally published at https://www.sertiscorp.com/
