A novel plug-and-play meta-black-box optimization module based on video streams for non-contact physiological signal extraction

Remote Photoplethysmography (rPPG) facilitates non-contact physiological monitoring yet faces substantial hurdles regarding motion artifacts and parameter optimization in complex, real-world scenarios. To address these limitations, this study proposes a novel plug-and-play framework DQPCA-TARE, with three key contributions: (1) A hemodynamics-based meta-black-box optimizer DQPCA (Double Q-Learning Pressure Circulation Algorithm) that adapts exploration–exploitation via Reynolds number and escapes local optima through dynamic fluid parameter tuning; (2) TARE (Temporal Adaptive Residual Enhancement) module that reconstructs long-range temporal correlations and refines features via frequency-domain decoupling and learnable residual fusion; (3) A label-free composite objective function integrating signal-to-noise ratio, spectral consistency, and stability constraints to guide optimization. Extensive evaluations demonstrate that DQPCA achieves an 89.66% winning rate on the CEC2017 benchmark, outperforming 11 state-of-the-art (SOTA) optimizers. On the PURE and MMPD datasets, DQPCA-TARE consistently improves the performance of 5 mainstream rPPG models across 4 diverse scenarios, delivering a maximum 85.87% reduction in mean absolute error (MAE). These results confirm that the proposed framework outperforms existing approaches in terms of robustness and generalization, particularly for challenging cross-dataset evaluations. The codes are available at: https://drive.mathworks.com/sharing/a010b1c4-c049-44f9-9322-6a55c54d8d08 . • Design TARE for long-range recovery via freq decoupling and residual fusion. • Propose DQPCA meta black-box optimizer with hemodynamics and Double Q-Learning. • Develop DQPCA-TARE plug-and-play, label-free objective for cross-dataset rPPG.