Restoring Morphological Features in rPPG Signals Using Generative Adversarial Networks with Time-Frequency Discriminator

Many vital signs such as heart rate, respiratory rate, and blood oxygen level can be obtained from remote photoplethysmography (rPPG) signals, making it a useful tool for non-invasive health monitoring. However, despite its potential capabilities, the rPPG signal still lacks the accuracy to detect some important features such as detailed waveform characteristics that are usually captured by PPG signal. In this paper, we present a machine learning-based rPPG signal reconstruction algorithm to address this drawback. The proposed method improves the accuracy of rPPG by restoring important features lost in the rPPG signal. We used Generative Adversarial Networks (GANs) to reconstruct the signal. We successfully restored both time domain features and frequency domain features using a dual discriminator. Cosine similarity, Pearson correlation coefficients, and dynamic time warping were used to evaluate the signals. The proposed method significantly improved the accuracy of rPPG signal compared with the true PPG signal extracted from videos.