Comparison of convolutional neural network architectures for face mask detection

Abstract

In 2020 World Health Organization (WHO) has declared that the Coronaviruses (COVID-19) pandemic is causing a worldwide health disaster. One of the most effective protections for reducing the spread of COVID-19 is by wearing a face mask in densely and close populated areas. In various countries, it has become mandatory to wear a face mask in public areas. The process of monitoring large numbers of individuals to comply with the new rule can be a challenging task. A costeffective method to monitor a large number of individuals to comply with this new law is through computer vision and Convolution Neural Network (CNN). This paper demonstrates the application of transfer learning on pre-trained CNN architectures namely; AlexNet, GoogleNet ResNet-18, ResNet-50, ResNet-101, to classify whether or not a person in the image is wearing a facemask. The number of training images are varied in order to compare the performance of these networks. It is found that AlexNet performed the worst and requires 400 training images to achieve Specificity, Accuracy, Precision, and F-score of more than 95%. Whereas, GoogleNet and Resnet can achieve the same level of performance with 10 times fewer number of training images.