ABSTRACT
A crucial task with diseases, such as COVID-19, is accurate forecasting of cases for early detection of spikes, which allows policymakers to adjust local restrictions. The use of face masks to prevent disease spread among the general population has become widespread due to the COVID-19 pandemic. While predictive models for COVID-19 case counts exist, capturing localized information about mask usage has the potential to improve prediction accuracy. In this paper, we develop time series models that utilize Twitter image data for COVID-19 case count prediction. A crucial part of such a model is the accurate detection of face mask presence in Twitter images, which we train a convolutional neural network (CNN) to perform. While multiple datasets exist to train CNNs for face mask detection, existing datasets do not adequately represent the complexity nor the diversity in social media images. To address this and create a sufficiently accurate CNN for use with social media images, we also present a new social media face mask image dataset designed for the training of CNNs to detect the presence of face masks in complex real-world images, such as social media images. The presented dataset consists of approximately 120k images and attempts to more adequately account for diversity in ethnicity, mask type, and physical orientation of individuals in images than existing datasets. We demonstrate the effectiveness of both the CNN model for face mask detection and the resulting time series model trained on data obtained from applying the CNN model to historical twitter data, illustrating that data on the presence of masks in social media images can increase predictive accuracy of time series models for COVID-19 case counts. © 2022 IEEE.