Identifying Image of the Correct Use of Face Mask Using Semantic Segmentation Technique

In recent years, the world is facing Covid-19 pandemic which has spread to more than 200 countries. WHO recommend everyone to always wearing a mask and keeping a distance to reduce transmission since Covid-19 is very susceptible to infection in a crowded area. In fact, many people misuse masks, such as wearing a mask but not covering the nose, and thus, monitoring the correct use of masks on a large-scale area is not easy. A technology implementing a high precision computer vision is needed to help monitoring the correct use of human mask automatically. This paper proposes a deep learning method that performs semantic segmentation and classification tasks to precisely identify the use of human face mask. Since it is rarely done so far, a sufficient dataset for this task is still lacking. Therefore, we also construct a dataset for face mask semantic segmentation task, including the fine-grained annotated ground truth. Based on our experiments, the proposed method that uses U-Net base model provides the best Mean IoU performance, which is 95%, compared to several comparative models. The segmentation output is then forwarded to a classification process, to decide whether it is a correct or an incorrect use of mask, and provides an accuracy rate that reaches 100%. Details of the experimental results are shown both quantitatively and qualitatively in this paper. The current results of this study may inspire the development of a better system in the future.