Detecting incorrect mask wearing using out-of-distribution detection

Abstract

Face mask detection has been a significant task since the Covid-19 pandemic began in early 2020. While various researches on mask-face detection techniques up to 2021 are available, only a few have been studied on the three classes (i.e., wearing mask, without mask, and incorrect mask-wearing). This is due to the difficulty in collecting and annotating images of incorrect mask-wearing. As a result, this class in the research has a lower detection accuracy than the other two classes. The objectives of this dissertation are focused on the two-fold: To provide a new dataset of mask faces from Wider Face and Kaggle; To propose a new framework named Out-of-distribution Mask (OOD-Mask) to perform the three-class detection task using only two-class training data. This is achieved by treating the incorrect mask-wearing situation as an anomaly class, leading to a reasonable performance in the absence of training data for the third class.