Abstract

The use of face masks to prevent disease spread among the general population has become widespread due to the COVID-19 pandemic. The ability to accurately detect and monitor the trends in face mask usage is crucial to understanding and predicting hotspot areas for both current and future pandemics. In this work, we investigate the detection of face masks in social media images using deep learning, specifically convolutional neural networks (CNNs). The use of CNNs for face mask detection has been explored by the research community; however, a common limiting factor has been the lack of a large, diverse image dataset of masked individuals. Current datasets are typically too small and limited in diversity or artificially created by superimposing medical masks onto faces. These approaches fail to reflect the diversity of real life images and unrealistically assume CNNs can always perfectly see the target's face. This project investigates the creation of a social-media-based face mask image dataset that reflects the scale needed for deep learning and the diversity (mask types, positions of people, and ethnicity) of real life. We have gathered approximately 120k images containing people tweeted from different cities. Mechanical Turk is used to label the images based on the presence of a face mask. Using this dataset, we train a CNN model to detect the presence of face masks in social media images and compare the results to existing approaches. We then deploy our model to detect trends in face mask usage in San Francisco over time.