Frequency-Preserving Readout Design and Error Characterization for CWT-Based rPPG Heart-Rate Estimation

Remote photoplethysmography (rPPG) estimates heart rate (HR) from facial color variations. In continuous wavelet transform (CWT) scalograms, HR is represented by position along the frequency axis, whereas global average pooling (GAP) does not retain this coordinate explicitly. We investigated whether a structured frequency-localizing readout improves HR estimation and analyzed the remaining errors. Leave-one-subject-out evaluations were conducted on three datasets (PURE, UBFC-rPPG, and BH-rPPG) using a fixed chrominance-based signal-extraction and CWT pipeline. The proposed model combined a frequency-preserving convolutional neural network backbone with a soft-argmax readout over HR coordinates. Using the same backbone and training protocol, soft-argmax reduced segment-pooled mean absolute error (MAE) relative to the matched GAP readout from 8.09 to 2.98 beats per minute on PURE and from 5.14 to 3.82 beats per minute on UBFC-rPPG, with significant subject-level improvements on both datasets. Residual analyses indicated that major errors could originate from color projection, non-cardiac spectral components, motion, and HR-range mismatch. On BH-rPPG, the proposed model achieved the lowest MAE under low illumination, while the evaluated CWT-based estimators showed smaller illumination-induced degradation than the evaluated fast Fourier transform (FFT)-based estimators. These findings support structured soft-argmax readout design and highlight the need to detect or mitigate degraded input signals.