Noise-resilient spatial-channel refinement for robust remote photoplethysmography.

Introduction: Remote photoplethysmography (rPPG) enables non-contact heart rate measurement from facial videos with growing health monitoring applications. Despite advances in deep learning methods, most existing approaches lack task-specific feature refinement across spatial and channel dimensions, which can dilute pulsatile information. Methods: We propose two key contributions to address this challenge. First, a Multi-scale Difference Fusion Stem (MDFS) captures temporal dynamics through learnable fusion of multi-scale frame differences, providing noise-resilient representations. Second, a Spatial-Channel Optimized Pulse Enhancement (SCOPE) module performs adaptive spatial-channel refinement with gated fusion to amplify pulse-related features while suppressing interference. Results: Comprehensive experiments on PURE, UBFC-rPPG, and MMPD datasets demonstrate state-of-the-art performance in both intra-dataset and cross-dataset evaluations. Discussion: These results indicate that explicitly constructing multi-scale temporal cues and performing controlled spatial-channel refinement are effective for robust rPPG estimation under low-SNR conditions.