Progressive hierarchical attention and frequency-domain decoupled physiological loss for unsupervised remote photoplethysmography.

Remote photoplethysmography (rPPG) enables contactless heart rate (HR) estimation from facial video. Unsupervised methods still face two coupled limitations: fixed backbones lack hierarchical feature refinement aligned with pulsatile signal formation, and contrastive objectives mainly regulate inter-sample spectral relations rather than the single-waveform morphology used for bandpass filtering and dominant-peak HR readout at test time.

Approach. We propose progressive hierarchical attention (PHA) and a frequency-domain decoupled physiological (FDP) loss on a single-branch three-dimensional CNN trained with Contrast-Phys-style power-spectral-density contrastive learning. PHA cascades spatial, temporal, and channel attention with learnable stage-wise fusion to suppress interference in the spatial-to-temporal-to-chrominance order of rPPG formation. FDP anchors each waveform at a dynamic fundamental frequency and applies out-of-band suppression with confidence-gated harmonic constraints.

Main results. On UBFC-rPPG, PURE, and COHFACE under in-domain and cross-dataset protocols, PHA and FDP yield complementary gains and remain competitive with contrastive and structure-decoupling baselines, while retaining a single-branch inference pipeline without background ROI.

Significance. By coupling feature-level hierarchical refinement with loss-level spectral anchoring, the framework aligns unsupervised training with practical frequency-domain HR readout and offers an interpretable, deployment-friendly paradigm for video-based physiological measurement.