rPPG-TFCL: Time-frequency consistency learning for robust remote physiological measurement

Remote photoplethysmography (rPPG) enables non-contact estimation of physiological signals from facial videos by capturing subtle skin color variations. However, existing frameworks fail to preserve two critical properties of physiological signals: temporal directionality, which reflects the asymmetric nature of systolic and diastolic phases, and waveform fidelity, which is essential for reliable cardiovascular monitoring. To address these challenges, we propose a unified Time-Frequency Consistency Learning framework, named rPPG-TFCL, which incorporates physiologically grounded priors at both the architectural and regularization levels. In particular, a Time-Reversal Consistency Strategy (TRCS) is proposed to enforce predictive symmetry between forward and reversed video sequences, thereby aligning the model with the inherent reversibility of cardiovascular dynamics. In parallel, a Hierarchical Spatio-Temporal Transformer (HSTT) is designed, in which the Discrete Wavelet Transform (DWT) is embedded into a multi-scale attention architecture to decouple low- and high-frequency components, enabling the model to capture subtle hemodynamic fluctuations across temporal and spectral resolutions. Furthermore, we introduce a time-reversal consistency loss to quantify reversibility in predictions and a frequency-domain consistency constraint based on Top-K spectral energy alignment. These strategies synergistically optimize the temporal and frequency-domain behavior of the model, resulting in improved prediction stability and enhanced spectral fidelity. Extensive experiments on three public datasets (UBFC-rPPG, PURE, and MMPD) demonstrate that the proposed method outperforms existing approaches in both periodicity modeling and high-fidelity waveform reconstruction, facilitating more robust and interpretable rPPG-based physiological monitoring.