摘要
Remote photoplethysmography (rPPG) enables non-contact physiological monitoring by capturing subtle skin color variations induced by cardiac cycles. Despite its promise, rPPG remains highly sensitive to environmental illumination. Existing illumination-aware methods suffer from two key limitations: (1) limited out-of-distribution (OOD) generalization due to domain discrepancies between training and testing data, and (2) reliance on the Lambertian assumption, which neglects non-Lambertian specular highlights that frequently corrupt facial skin signals. In this paper, we propose SHINE-PPG (Specular-Highlight Intrinsic Network for rPPG Estimation), a novel framework that leverages non-Lambertian intrinsic decomposition to decouple facial videos into illumination, reflectance, and specular components in a self-supervised manner. By isolating physiological information within the intrinsic reflectance, our method effectively suppresses both ambient lighting variations and surface highlights to recover high-fidelity rPPG signals. To further enhance robustness, we introduce an adversarial illumination enhancement strategy that dynamically synthesizes challenging unseen lighting conditions during training, significantly improving OOD generalization. Extensive experiments on five benchmark datasets demonstrate that SHINE-PPG consistently outperforms previous methods, particularly under complex and dynamic illumination scenarios. The code is available at https://github.com/Edmond-Yang/SHINE-PPG.