Noise-resilient spatial-channel refinement for robust remote photoplethysmography
Xianwei Zhang1, Feng Qiao1, Qiaochu Zang1
1School of Information Science and Engineering, Shandong University, Qingdao, China.
Frontiers in Bioengineering and Biotechnology
|May 25, 2026
Summary
This study introduces novel deep learning methods for non-contact remote photoplethysmography (rPPG) to improve heart rate estimation. The new techniques enhance signal quality for more accurate health monitoring from facial videos.
Area of Science:
- Biomedical Engineering
- Computer Vision
Background:
- Remote photoplethysmography (rPPG) offers non-contact heart rate measurement from facial videos, crucial for health monitoring.
- Current deep learning methods often struggle with feature refinement, diluting vital pulsatile information and impacting accuracy.
Purpose of the Study:
- To develop advanced deep learning techniques for robust rPPG signal extraction.
- To enhance the accuracy of non-contact heart rate estimation, especially in low signal-to-noise ratio (SNR) conditions.
Main Methods:
- Proposed a Multi-scale Difference Fusion Stem (MDFS) to capture temporal dynamics using fused multi-scale frame differences for noise resilience.
- Introduced a Spatial-Channel Optimized Pulse Enhancement (SCOPE) module for adaptive feature refinement, amplifying pulse signals and reducing interference.
Main Results:
- Achieved state-of-the-art performance on PURE, UBFC-rPPG, and MMPD datasets.
- Demonstrated superior results in both intra-dataset and cross-dataset evaluations, validating the model's robustness.
Conclusions:
- Explicitly constructing multi-scale temporal cues is effective for rPPG estimation.
- Controlled spatial-channel refinement significantly improves robustness in low-SNR environments, enhancing remote heart rate monitoring accuracy.