Evaluating Deep Transfer Learning Models for Detecting Various Face Mask Wearings

In Malaysia, wearing a face mask is no longer mandatory to prevent COVID-19. However, such wearing may be important in future if another outbreak occurs. Besides, wearing a face mask is also important in environments such as cleanrooms, operating theatres and crowded places. For these reasons, developing automated systems for detecting face mask-wearing is crucial. This study evaluates deep transfer learning models to detect four categories of face mask wearing: (i) not wearing masks, (ii) wearing single masks, (iii) wearing masks incorrectly, and (iv) wearing double masks. Transfer learning methods were adopted by using five pre-trained models: (i) VGG-16, (ii) MobileNetV2, (iii) ResNet-152, (iv) InceptionV3 and (v) Xception. These models were trained based on 2,000 images collected from various sources and then augmented. The results showed that ResNet-152 outperformed the others by achieving 86.67% accuracy on the testing set (120 images from the other distribution) and 84.47% accuracy on the videos captured using a smartphone.