Scholarlink

This study evaluates the application of Gradient-weighted Class Activation Mapping (Grad-CAM) to identify key image regions for enhancing pulse wave estimation using traditional signal analysis methods. The approach assumes that Grad-CAM masking enables the selection of relevant image areas, improving the Signal to Noise Ratio (SNR). Experiments were conducted on the PURE dataset using the TS-CAN model and rPPG-Toolbox. Grad-CAM maps identified facial regions most influential for model prediction, allowing the exclusion of areas not contributing to accurate estimation. The study also explored different temporal window sizes and their impact on signal quality. Method evaluation included SNR \(_{raw}\) , SNR \(_{max}\) , SNR \(_{sum}\) , and Hjorth descriptors. Results confirmed that Grad-CAM masking enhances rPPG signal quality, enabling more precise heart rate estimation. Statistical analysis validated the significance of the findings, highlighting the potential of interpretable deep learning methods in signal analysis.

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

Explainable Estimation of Blood Volume Pulse Signals from Video Sequences Using a Combination of Deep Learning Models and Signal Processing Methods

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Milena Sobotka,
  • Kamil Kopryk,
  • Muhammad Usman,
  • Jacek Rumiński

摘要

This study evaluates the application of Gradient-weighted Class Activation Mapping (Grad-CAM) to identify key image regions for enhancing pulse wave estimation using traditional signal analysis methods. The approach assumes that Grad-CAM masking enables the selection of relevant image areas, improving the Signal to Noise Ratio (SNR). Experiments were conducted on the PURE dataset using the TS-CAN model and rPPG-Toolbox. Grad-CAM maps identified facial regions most influential for model prediction, allowing the exclusion of areas not contributing to accurate estimation. The study also explored different temporal window sizes and their impact on signal quality. Method evaluation included SNR \(_{raw}\) , SNR \(_{max}\) , SNR \(_{sum}\) , and Hjorth descriptors. Results confirmed that Grad-CAM masking enhances rPPG signal quality, enabling more precise heart rate estimation. Statistical analysis validated the significance of the findings, highlighting the potential of interpretable deep learning methods in signal analysis.