Abstract
With COVID-19 spreading all over the world and restricting our daily lives, the use of face masks has become very important, as it is an efficient way of slowing down the spread of the virus and an important piece to continue our daily tasks until vaccination is completed. People have been fighting this disease for a long time, and they are bored with the precautions, so they act carelessly. In this case, automatic detection systems are very important to keep the situation under control. In this research, deep learning models are trained with as little input data as possible in order to obtain an accurate face mask-wearing condition classification. These classes are mask-correct, mask wrong, and no mask, which refers to proper face mask use, improper face mask use, and no mask use, respectively. DenseNets, EfficientNets, InceptionResNetV2, InceptionV3, MobileNets, NasNets, ResNets, VGG16, VGG19, and Xception are the networks used in this study. The highest accuracy was obtained by the InceptionResNetV2 and Xception networks, with 99,6%. When other performance parameters are taken into consideration, the Xception network is a step forward. VGG16 and VGG19 also show an accuracy rate over 99%, with 99,1 and 99,4%, respectively. These two networks also had higher FPS and the two lowest initialization times during implementation. A comparison with recent studies was also carried out to evaluate the obtained accuracy. It was found that a higher accuracy can be obtained with the possible minimum input size.
Publisher
Universidad Nacional de Colombia
Subject
General Engineering,Building and Construction