Determining the Reliability of Personal Masks with Convolutional Neural Networks

Abstract

During the COVID-19 pandemic, which is a worldwide disaster, it has been proven that one of the most important methods to struggle the transmission of such diseases is the use of face masks. Due to this pandemic, the use of masks has become mandatory in Turkey and in many other countries. Since some surgical masks do not comply with the standards, their protective properties are low. The aim of this study is to determine the reliability of personal masks with Convolutional Neural Networks (CNNs). For this purpose, first, a mask data set consisting of 2424 images was created. Subsequently, deep learning and convolutional neural networks were employed to differentiate between meltblown surgical masks and non-meltblown surgical masks without protective features. The masks under investigation in this study are divided into 5 classes: fabric mask, meltblown surgical mask, meltblown surgical mask, respiratory protective mask and valve mask. Classification of these mask images was carried out using various models, including 4-Layer CNN, 8-Layer CNN, ResNet-50, DenseNet-121, EfficientNet-B3, VGG-16, MobileNet, NasNetMobile, and Xception. The highest accuracy, 98%, was achieved with the Xception network.