rPPG Signal Estimation: Self-Supervised Contrastive Learning Based on Frequency-Noise Data Augmentation

Remote Photoplethysmography (rPPG) is a typical method for measuring physiological indicators remotely. Based on the principle that the absorption of light by the human skin changes with the flow of subcutaneous blood, rPPG technology captures the periodic changes in skin color through sensors such as cameras, and extracts blood volume pulse signals (also known as rPPG signals) to measure physiological indicators related to the cardiac cycle, such as heart rate, respiratory rate, and heart rate variability. In the rPPG signal estimation task, annotated video datasets are scarce. Facial Video-based Remote Physiological Measurement via Self-supervised Learning proposes a frequency augmentation self-supervised contrastive learning framework that not only considers extracting rPPG signals from unlabeled facial videos, but also considers situations where the sample has a different signal frequency from the original sample. However, data quality is a concern, such as in the data collection environment and transmission process, which may cause unnecessary noise in the data. Therefore, based on the framework of Facial Video-based Remote Physiological Measurement via Self-supervised Learning, this paper introduces noise augmentation in the data augmentation stage, simulating several common data noises, allowing the model to consider both frequency and noise effects, and enhancing the model's generalization ability to different data. Through experiments, it has been shown that the model proposed in this paper has a slight reduction in the error of extracting rPPG signals when facing noisy data compared to the original model.