Intelligent system based on a convolutional neural network for identifying people without breathing masks

Abstract

The COVID-19 pandemic is having a huge impact on people and communities. Many organizations face significant disruptions and
issues that require immediate response and resolution. Social distancing, breathing masks and eye protection as preventive measures
against the spread of COVID-19 in the absence of an effective antiviral vaccine play an important role. Banning unmasked shopping
in supermarkets and shopping malls is mandatory in most countries. However, with a large number of buyers, the security is not able
to check the presence of breathing masks on everyone. It is necessary to introduce intelligent automation tools to help the work of
security. In this regard, the paper proposes an up-to-date solution – an intelligent system for identifying people without breathing
masks. The proposed intelligent system works in conjunction with a video surveillance system. A video surveillance system has a
structure that includes video cameras, recorders (hard disk drives) and monitors. Video cameras shoot sales areas and transmit the
video image to recording devices, which, in turn, record what is happening and display the video from the cameras directly on the
monitor. The main idea of the proposed solution is the use of an intelligent system for classifying images periodically received from
cameras of a video surveillance system. The developed classifier divides the image stream into two classes. The first class is “a person in a breathing mask” and the second is “a person without a breathing mask”. When an image of the second class appears, that is,
a person who has removed a breathing mask or entered a supermarket without a breathing mask, the security service will immediately receive a message indicating the problem area. The intelligent system for image classification is based on a convolution neural
network VGG-16. In practice, this architecture shows good results in the classification of images with great similarity. To train the
neural network model, the Google Colab cloud service was used – this is a free service based on Jupyter Notebook. The trained model is based on an open source machine learning platform TensorFlow. The effectiveness of the proposed solution is confirmed by the
correct processing of the practically obtained dataset. The classification accuracy is up to 90 %.