LightweightPhys: A Lightweight and Robust Network for Remote Photoplethysmography Signal Extraction

Remote Photoplethysmography (rPPG) has emerged as a promising technology for non-contact heart rate monitoring by analyzing subtle color variations in facial videos, which are caused by changes in blood volume. While traditional rPPG methods often struggle with environmental noise and motion artifacts, deep learning-based approaches have shown improved accuracy but typically come with high computational costs, limiting their applicability in real-time, resource-constrained scenarios. To address these challenges, this paper introduces LightweightPhys, a novel 3D-CNN-based network that combines Depthwise Separable Convolution (DSC) and a newly proposed Simulated Temporal Noise Suppression (Sim-TN) module. The DSC module significantly reduces the computational complexity of the model by decoupling spatial and channel-wise convolutions, making it more efficient for deployment on low-power devices. Meanwhile, the Sim-TN module enhances the robustness of feature extraction by effectively suppressing temporal noise, which is a common issue in rPPG signal processing due to environmental factors and subject movement. Extensive evaluations on two widely-used datasets, PURE and UBFC-rPPG, demonstrate that LightweightPhys achieves state-of-the-art performance in heart rate estimation while maintaining significantly lower computational overhead compared to existing deep learning models. This makes LightweightPhys particularly suitable for real-time health monitoring applications on resource-constrained devices, such as wearable gadgets and mobile health platforms.