Neural network based face mask detection using thermal imaging sensors

The current COVID situation has led to various autonomous systems checking for correctly attached face masks. Camera systems which are trained on a large number of examples show an impressive performance in the recognition task. Nevertheless they require a visible image which is often a concern for data privacy issues. Therefore we are proposing a face mask detection system which is based on a thermal imaging camera in the long wave infrared spectrum from 8 to 14 micrometer. These image sensors preserve data privacy and have a moderate resolution and frame rate which is very well suited for the task. We have implemented two different approaches with a binary classifier on full images showing a single person and a face localisation system in combination with a binary classifier. The face localisation system allows the usage of the detector in multi-person scenarios as public sites. The first approach is based on a self-designed and trained convolutional neural network whereas the second approach uses a pre-trained and adapted YOLOv5s network. Both approaches have reached a high recognition accuracy on our self-created dataset and are successfully demonstrated in public environments.