Bi-level weighted mixed-domain self-attention network for non-contact heart rate estimation

Remote Photoplethysmography (rPPG) technique plays a pivotal role in non-contact heart rate (HR) estimation. The existing deep learning-based technology has obtained better performance than the traditional method. However, there is a large amount of redundant information in the pulse signal due to noise from light, motion, etc. Meanwhile, single region of interest (ROI)-based methods cannot get accurate prediction results when the region is occluded. For this reason, this paper proposes a Bi-level weighted mixed-domain self-attention network (Bi-WMSN) for non-contact HR estimation. In the first-level weighting, ROI adaptive weighted blocks (RAWB) are developed to adaptively fit the weights of signals from different ROI regions. In the second-level weighting, a dual-channel mixed-domain adaptive weighted network (DMAWN) is designed to adaptively suppress redundant information in the input signal and highlight features related to HR estimation. The weights of different feature information are calibrated to improve the performance of HR estimation. Finally, the results of the Bi-level weighting are input into the HR estimation framework to obtain the estimated HR. The proposed method outperforms those state-of-the-art in terms of many evaluation metrics on the two datasets: MAHNOB-HCI (MAE = 2.18, RMSE = 3.17), PURE (MAE = 2.08, RMSE = 3.18).