A comparative study of VGG models performance for masked face recognition

Officials in public health have found that using a mask for the face can minimize the spread of COVID-19. This disease may mutate into more dangerous forms that threaten human life, which requires wearing a face mask to limit its spread. Thus, people in enclosed locations need to use face masks to verify their identity on their laptops or mobile phones. Recognizing the identity of masked faces is a challenging problem for computer vision models since the features required to accurately predict the identity of an individual are reduced from the whole face to just the eye and sometimes the forehead. This study is built on existing pre-trained VGG16 and VGG19 architectures to solve the problem of recognizing a person's identity when wearing a face mask. A new method is proposed using Otsu's thresholding and one of the morphological operations called opening to separate the input masked image into multiple interest regions then YOLOv5 network is applied to classify these regions into masked and unmasked. In the experimental results, the accuracy was evaluated for the VGG models. Promising results are obtained for the VGG19 model using the HFRD dataset. A comparison is made with VGG16 results, VGG19 satisfied good accuracy but was also heavier and required more memory and computational time.