Scholarlink

<p>Remote photoplethysmography (rPPG) enables noncontact measurement of physiological signals from facial video, with applications ranging from health monitoring to affective computing. Existing deep-learning approaches–including convolutional neural network (CNN) based pipelines and Transformer architectures–either struggle to capture the long-range periodic patterns intrinsic to rPPG or require large annotated datasets and heavy computation. We introduce a multi-scale periodic attention block that hierarchically captures cardiac cycles at multiple temporal resolutions and injects a physiological-prior bias–a Gaussian-shaped temporal offset penalty peaked at plausible inter-beat intervals–into the sparse attention scores, and a Raw-Video Masked Autoencoder (M-AE) block that operates directly on unprocessed face frames. In this M-AE, we randomly mask spatio-temporal cuboids in the raw video stream and train the network to reconstruct them under a periodicity-aware loss (combining pixel Mean Squared Error (MSE) with a Pearson correlation coefficient (<InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11760_2025_4493_Article_IEq1.gif" Format="GIF" Height="12" Rendition="HTML" Resolution="72" Type="Linedraw" Width="13" /> </InlineMediaObject> <EquationSource Format="TEX">\(\rho \)</EquationSource> <EquationSource Format="MATHML"><math> <mi>ρ</mi> </math></EquationSource> </InlineEquation>) term on the recovered waveform), thereby encouraging the model to discover subtle heartbeat-driven color changes end-to-end–no Spatio-Temporal Map (STMap) preprocessing required. Our periodicity module adaptively attends to both short- and long-term cardiac cycles, while Raw-Video M-AE reconstruction fosters robust feature learning in low-annotation regimes. Evaluated on the PURE and UBFC-rPPG benchmarks, our method achieves up to 25.8% lower Mean Absolute Error (MAE) and 12.4% higher Signal-to-Noise Ratio (SNR) compared to a strong baseline model. These results suggest that the proposed framework offers a promising trade-off between accuracy and efficiency, making it suitable for real-time remote physiological measurement with reduced reliance on labeled data.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

SRMFormer: A Physiological-Prior Multi-Scale Periodic Attention Transformer with Raw-Video MAE for Remote Physiological Measurement

  • 期刊论文

  • 发表日期：

    2025年7月9日

查看全文
  • Libin Zhang,
  • Tao Peng,
  • Xiaolu Li,
  • Hanguang Xiao,
  • Xuewei Chen

摘要

Remote photoplethysmography (rPPG) enables noncontact measurement of physiological signals from facial video, with applications ranging from health monitoring to affective computing. Existing deep-learning approaches–including convolutional neural network (CNN) based pipelines and Transformer architectures–either struggle to capture the long-range periodic patterns intrinsic to rPPG or require large annotated datasets and heavy computation. We introduce a multi-scale periodic attention block that hierarchically captures cardiac cycles at multiple temporal resolutions and injects a physiological-prior bias–a Gaussian-shaped temporal offset penalty peaked at plausible inter-beat intervals–into the sparse attention scores, and a Raw-Video Masked Autoencoder (M-AE) block that operates directly on unprocessed face frames. In this M-AE, we randomly mask spatio-temporal cuboids in the raw video stream and train the network to reconstruct them under a periodicity-aware loss (combining pixel Mean Squared Error (MSE) with a Pearson correlation coefficient ( \(\rho \) ρ ) term on the recovered waveform), thereby encouraging the model to discover subtle heartbeat-driven color changes end-to-end–no Spatio-Temporal Map (STMap) preprocessing required. Our periodicity module adaptively attends to both short- and long-term cardiac cycles, while Raw-Video M-AE reconstruction fosters robust feature learning in low-annotation regimes. Evaluated on the PURE and UBFC-rPPG benchmarks, our method achieves up to 25.8% lower Mean Absolute Error (MAE) and 12.4% higher Signal-to-Noise Ratio (SNR) compared to a strong baseline model. These results suggest that the proposed framework offers a promising trade-off between accuracy and efficiency, making it suitable for real-time remote physiological measurement with reduced reliance on labeled data.