DSE-NN: Deeply Supervised Efficient Neural Network for Real-Time Remote Photoplethysmography.

Area of Science:

• Biomedical Engineering
• Computer Science
• Signal Processing

Background:

• Non-contact remote photoplethysmography (rPPG) measures vital signs unobtrusively.
• Existing deep learning rPPG models lack interpretability, functioning as black boxes.
• Need for interpretable and efficient deep learning models in rPPG analysis.

Purpose of the Study:

• To develop an interpretable and lightweight deep learning model for rPPG.
• To visualize temporal and spectral representations in hidden layers of rPPG networks.
• To improve training and inference speeds while maintaining high performance.

Main Methods:

• Proposed a method for visualizing temporal and spectral representations in hidden layers.
• Implemented deep supervision on spectral representations of intermediate layers.
• Optimized the network for lightweight architecture, improving training and inference efficiency.
• Conducted thorough ablation studies on public datasets.

Main Results:

• Achieved high accuracy in vital sign measurement, with an RMSE of 1 bpm on the PURE dataset.
• Outperformed state-of-the-art methods on the V4V dataset with an RMSE of 6.65 bpm.
• Demonstrated fast convergence from the first epoch, indicating improved learning efficiency.
• The proposed spectral deep supervision acts as a regularizer, enhancing convergence speed.

Conclusions:

• The developed method enhances interpretability and efficiency in deep learning-based rPPG.
• The model achieves state-of-the-art performance in vital sign estimation.
• The approach shows potential for general application in spectral domain learning and periodicity-based regression tasks.