Mask and respirator detection: analysis and potential solutions for a frequently ill-conditioned problem

Abstract

During the coronavirus pandemic, the mask detection problem has become of particular interest. Usually, the goal is to create a system that can detect whether or not a person is wearing a mask or respirator. However, this tends to trivialize a problem that hides a greater complexity. In fact, people wear masks or respirators in various ways, many of which are incorrect. This makes the problem ill-conditioned and creates a bias compared to training cases, with the consequence that these systems have a considerably lower accuracy when used in practice. We claim that focusing on the ways in which a mask can be worn and classifying the problem not as binary but at least as ternary, thus adding an intermediate class containing all those ways in which a mask or respirator can be worn incorrectly, could help address this problem. For this reason, this paper describes and puts to the proof the Ways to Wear a Mask or a Respirator Database (WWMR-DB). It has a fine classification of the most common ways in which a mask or respirator is worn, which can be used to test how mask detection systems work in cases that resemble the real ones more. It was used to test a neural network, the ResNet-152, which was trained on less fine databases, like the Face-Mask Label Dataset and the MaskedFace-Net. The mixed results denote the shortcomings of these databases and the need to enhance them or resort to finer databases.