UL-Phys:Ultra-lightweight remote physiological measurement in facial videos based on unsupervised learning

Remote photoplethysmography (rPPG) enables non-contact monitoring of vital signs using facial videos, but current supervised learning methods often rely on complex architectures and large annotated datasets, limiting their practicality in real-time and resource-constrained scenarios. This paper addresses these limitations by proposing UL-Phys, an ultra-lightweight self-supervised framework for rPPG signal estimation. From a research standpoint, we reformulate the rPPG task as a linear self-supervised reconstruction problem, introducing a novel frequency-constrained objective to extract inherent periodic information without requiring ground truth labels. The framework integrates a lightweight 3D spatiotemporal encoder-decoder network, and a neuroscience-inspired hybrid attention module to enhance pulsatile signal regions while suppressing noise. Experimental evaluations on PURE and UBFC-rPPG datasets demonstrate that UL-Phys achieves superior performance compared to existing supervised and self-supervised baselines, while significantly reducing model complexity and inference latency. Our method also shows strong generalization across datasets, highlighting the value of embedding physiological priors into lightweight, self-supervised architectures. These findings offer a promising direction for scalable and deployable rPPG systems in real-world settings.