Comparative study of Face Mask Recognition using Deep Learning and Machine learning classifiers

Abstract

With around 194 million cases and around 4 million reported deaths, affecting 220 countries [1], Coronavirus (COVID-19) is still prevalent. Wearing facemasks in crowded areas is one of the undemanding and effective measures among the multitude of preventive guidelines provided by the World Health Organization (WHO). However, unruly humans are present; monitoring if people are wearing facemasks in dense areas is taxing and cumbersome. In this paper, we have experimented two ways of tackling facemask detection for comparison purposes: (1) by using transfer learning on four pretrained State-Of- The-Art (SOTA) models - Inception-V3, Resnet-50, VGG-16, and Densenet-121, (2) using these SOTA models as feature extractors and training ML classifiers (Support Vector Machine (SVM), Decision Tree, and Gaussian Naive Bayes) on them. Simulated Face Mask Dataset (SMFD) is used to train and validate all of the models, including data augmentation to enhance data samples. The SOTA models displayed exceptional validation accuracy (greater than 90%), with VGG-16 and ResNet-50 performing the best. Similarly, all combinations of SOTA-ML models have remarkable performance with the Densenet-121-SVM model obtaining highest accuracy with lesser training time.