Remote photoplethysmography signals enhancement based on generative adversarial networks

Remote photoplethysmography (rPPG) is a non-contact technology for measuring human physiological indicators. The rPPG signal is extracted through the periodic changes of the facial skin in the facial video. The rPPG signal extracted by most existing rPPG methods has low quality, and it only can be used for average heart rate (HR) and respiratory rate (RR), etc. But for the field of heart rate variability (HRV) analysis, higher quality rPPG signals are needed. This paper introduces a novel and efficient generative adversarial network method defined as rPPGGAN, which denoises and enhances the original rPPG signal extracted by a deep learning rPPG method to make it closer to the real rPPG signal. The HRV indicators were verified on the public datasets UBFC-rPPG and PURE, and both the time-domain indicators and frequency-domain indicators have been effectively improved. The method proposed in this paper is lightweight, efficient, and easy to integrate. It can effectively improve the rPPG signal quality after being integrated with other rPPG methods.