Recovering Pulse Waves From Video Using Deep Unrolling and Deep Equilibrium Models.

Area of Science:

• Biomedical Engineering
• Computer Vision
• Physiological Monitoring

Background:

• Camera-based contactless monitoring of vital signs, or imaging photoplethysmography (iPPG), is used in various applications.
• Existing iPPG methods often rely on model-based priors or end-to-end deep learning.
• Accurate estimation of pulse rate and pulse rate variability from facial video remains a challenge.

Purpose of the Study:

• To introduce novel methods combining signal processing and deep learning within an inverse problem framework for iPPG.
• To estimate the underlying pulse signal, pulse rate, and pulse rate variability from facial video.
• To develop efficient deep learning models for iPPG signal denoising and vital sign inference.

Main Methods:

• Utilizing an inverse problem framework that integrates signal processing and deep learning.
• Employing deep-network-based denoising operators through deep algorithm unfolding and deep equilibrium models.
• Estimating pulse signal, pulse rate, and pulse rate variability from facial video data.

Main Results:

• The proposed methods effectively denoise acquired facial signals.
• Accurate inference of underlying pulse rate and pulse rate variability is achieved.
• Pulse rate estimation performance is consistent with state-of-the-art methods on benchmarks.
• The methods achieve this with significantly fewer learnable parameters (less than one-fifth) compared to competing approaches.

Conclusions:

• The developed methods offer a robust and efficient approach to contactless vital sign monitoring using iPPG.
• Combining signal processing with deep learning in an inverse problem framework enhances iPPG accuracy and parameter efficiency.
• These findings advance the field of non-invasive physiological monitoring through advanced computational techniques.