Convolutional Neural Networks for Semantic Segmentation as a Tool for Multiclass Face Analysis in Thermal Infrared.

Area of Science:

• Computer Vision
• Medical Imaging
• Thermography

Background:

• Thermal infrared imaging offers unique physiological data.
• Accurate segmentation of facial regions is crucial for thermal analysis.
• Existing image-to-image translation methods provide a foundation for pixel-level classification.

Purpose of the Study:

• To adapt established convolutional neural network architectures for multiclass segmentation of thermal infrared face images.
• To develop a system for real-time, pixel-accurate temperature analysis of facial areas.
• To demonstrate the potential for applications such as infection detection (e.g., COVID-19).

Main Methods:

• Utilized convolutional neural networks (CNNs) for image-to-image translation.
• Created pixel-accurate class annotations for training data.
• Trained networks on a custom thermal infrared face database.
• Employed the intersection over union (IoU) metric for quantitative evaluation.

Main Results:

• Established CNN architectures were successfully trained for multiclass thermal face segmentation.
• The trained network accurately segmented unknown infrared face images into predefined classes.
• Demonstrated real-time face classification and relative temperature display from learned areas.
• Achieved high accuracy in quantitative evaluation using IoU.

Conclusions:

• The proposed methodology enables accurate, pixel-level thermal face analysis.
• The approach effectively focuses on relevant facial regions, excluding accessories.
• The technique is transferable to other quantitative thermography tasks, including materials characterization and quality control.