Face with Mask Detection in Thermal Images Using Deep Neural Networks

Area of Science:

• Computer Vision
• Artificial Intelligence
• Biomedical Imaging

Background:

• Facial detection is increasingly important, particularly during pandemics.
• Thermal imaging offers advantages like environmental resistance and temperature measurement capabilities.
• Existing facial detection methods need adaptation for thermal images, especially with face masks.

Purpose of the Study:

• To evaluate deep learning algorithms for face detection in thermal images.
• To assess performance on faces covered by virus protective masks.
• To investigate the impact of preprocessing techniques on detection accuracy.

Main Methods:

• A dataset of over 7900 thermal face images (with and without masks) was created.
• Deep learning models, including Yolov3, were trained and evaluated.
• Transfer learning from visible light images was employed.
• Data preprocessing methods were analyzed for their influence on results.

Main Results:

• Transfer learning yielded a mean average precision (mAP) above 82% for many models.
• The Yolov3-based model achieved an mAP of at least 99.3% and a precision of 66.1%.
• Inference times were suitable for deployment on low-cost platforms.

Conclusions:

• Deep learning, particularly Yolov3, demonstrates high effectiveness for face detection in thermal images, even with masks.
• The approach is viable for real-time public health applications.
• Further research into preprocessing can optimize performance.