Deep learning-based image enhancement for robust remote photoplethysmography in various illumination scenarios

Remote photoplethysmography (rPPG) is a non-invasive and convenient approach for measuring human vital signs using a camera. However, accurate measurement can be challenging due to the different illumination of the surrounding environment. In this study, we present a deep learning-based image enhancement model (IEM) inspired by the Retinex theory to improve the robustness of rPPG signal extraction and heart rate (HR) estimation in different lighting conditions. We fine-tuned the IEM with a time-shifted negative Pearson correlation between the PPG signal ground truth and the predicted rPPG signal from a pre-trained 3D CNN (PhysNet). We evaluated our method using a publicly available dataset (BH-rPPG) of various lighting scenarios and our own private dataset. Our results demonstrate that our proposed model is generalizable and significantly improves rPPG extraction and HR estimation accuracies across a range of illumination intensities.