Adaptive-Weight Network for Imaging Photoplethysmography Signal Extraction and Heart Rate Estimation

Recently, imaging photoplethysmography (IPPG), a noncontact heart rate (HR) measurement based on face video, has attracted extensive attention. However, due to various noise interferences, such as illumination variation and motion, the IPPG signal extracted from a single fixed region of interest (ROI) is sometimes unstable to estimate the accurate HR. To extract the stable and high-quality IPPG signal for accurate HR estimation, we propose a two-stage network called adaptive-weight network, which can fully utilize the information from multiple ROIs in this work. The IPPG weight network as the first block adaptively calculates the weights for the IPPG signals from multiple ROIs. By combining the raw IPPG signals with the assigned weights, the optimized IPPG signal with higher quality can be obtained. In the second block, the HR estimation network with a long short-term memory (LSTM) model is deployed to map the optimized signal to the HR. As a verification, we test the proposed method on the MAHNOB-HCI dataset with a standard deviation (STD) of 7.55 and a root mean square error (RMSE) of 7.65 in estimating the HR, which could outperform some other typical IPPG methods. Furthermore, the optimized IPPG signal can be used for other physiological parameter measurements [e.g., HR variability (HRV)].