An ensemble and deep transfer learning approach for face mask detection to control spread of COVID-19.

Abstract

The introduction of Covid-19 in 2019 and the recommendations of healthcare regulators for wearing a facemask in public created a need for an automated facemask identification system. As the spread of Covid-19 can be greatly controlled by following SOPs, one of which is wearing a facemask. Deep learning has demonstrated impressive performance in visual recognition tasks, and we intended to use machine learning and deep learning algorithms to solve the problem of automated facemask identification. Because of the problem's novelty, there is a scarcity of a large and representative dataset for facemask identification. As a result, we selected two public datasets, one with 7,553 images and the other with 11,792 images. Both of these datasets contain images labelled with the classes "wearing a mask" and "not wearing a mask." After preparing the datasets in a suitable format, we trained four classification models on them. Transfer learning is used to train three of these classification models: Inception-V3, ResNet-50, and VGG19. The fourth model is an ensemble classifier that uses traditional machine learning algorithms (SVM, Nave Bayes, Logistic Regression, and Random Forest). To ensure a fair comparison, training and testing are conducted with a 70-30 split, and the same split is used in all four approaches. Inception-V3 outperforms the other four classifiers, with classification accuracies of 93% and 97% for Kaggle Facemask Detection Dataset and Prajna Bhandary’s Dataset, respectively.