Eye Recognition by YOLO for Inner Canthus Temperature Detection in the Elderly Using a Transfer Learning Approach

Area of Science:

• Biomedical Engineering
• Computer Vision
• Gerontology

Background:

• Early detection of illness in seniors is crucial for timely healthcare interventions.
• Elevated body temperature is a key indicator of viral infections.

Purpose of the Study:

• To develop and implement a multi-age thermal faces dataset for training object detection models.
• To create an eye detection model specifically for the elderly to accurately measure temperature.
• To compare the performance of different You Only Look Once (YOLO) models for thermal face analysis.

Main Methods:

• Preparation of a multi-age thermal faces dataset.
• Training of YOLOv5, YOLOv6, and YOLOv7 object detection models for eye detection.
• Application of transfer learning from a general YOLOv7 model to an elderly-specific YOLOv7 model.
• Analysis of model speed, accuracy (mAP@.5), and size.

Main Results:

• The YOLOv7 model demonstrated superior performance with a Mean Average Precision (mAP@.5) of 0.996 and 150 Frames Per Second (FPS).
• Eye detection enabled accurate temperature scanning, achieving a normalized error distance of 0.03.
• The transfer learning approach enhanced the elderly-specific YOLOv7 model's effectiveness.

Conclusions:

• A fast and reliable deep learning-based temperature detection model was successfully developed using non-contact infrared cameras.
• The specialized eye detection model offers a promising tool for early disease detection in the elderly population.