| 摘要: |
| 为了提高生成远程光电容积脉搏波描记法(remote Photoplethysmography，rPPG)信号的波形规律性，降低心率计算难度，提出了一种标准化rPPG的信号生成方法。将人脸视频输入到生成对抗网络的生成器中，生成器通过有标签数据进行监督训练，预测得到人脸视频中蕴含的rPPG信号。为了进一步优化生成器的预测结果，使用数学建模方法生成标准的rPPG信号，并将其与生成器预测的信号同时输入至判别器中进行对抗学习，不断优化生成器参数，使得生成器能够学习标准信号的波形分布。这样，生成器预测的信号波形在形态上更接近于真实rPPG信号的波形分布，从而有利于后续心率计算。在不同数据集上进行的实验结果表明，该方法可以显著提高预测的准确性，且拥有更高的信噪比。 |
| 关键词: 非接触式心率测量 远程光电容积脉搏波描记法 生成式对抗网络 |
| DOI：10.20079/j.issn.1001-893x.230512001 |
|
| 基金项目:中国博士后科学基金（2022M722248）；四川省无人系统智能感知控制技术工程实验室开放课题（WRXT2021-001） |
|
| Standard rPPG Signal Generation Based on Generative Adversarial Learning |
| TU Xiaoguang,HU Zhehao,HU Junqiang,LIU Jianhua,LEI Xia,LIU Yuang,WANG Yu,FENG Ziliang |
| (1.College of Computer Science,Sichuan University,Chengdu 610065,China;2.Institute of Electronic and Electrical Engineering,Civil Aviation Flight University of China,Guanghan 618307,China;3.Civil Aviation Flight University of China Xinjin Flight College,Chengdu 611431,China) |
| Abstract: |
| To improve the regularity of the generated remote photoplethysmography(rPPG) signal waveform and reduce the difficulty of heart rate calculation,a standardized rPPG signal generation method is proposed.Specifically,facial videos are input into the generator of a generative adversarial network(GAN),and the generator is supervised trained using labeled data to predict the underlying rPPG signal contained in the facial videos.To further optimize the generator搒 predictions,mathematical modeling is employed to generate standardized rPPG signals,which are then input along with the generator搒 predicted signals into a discriminator for adversarial learning.This iterative process continuously improves the generator parameters,enabling the generator to learn the waveform distribution of the standardized signals.Consequently,the waveform predicted by the generator is more similar in morphology to the distribution of real rPPG signals,thereby facilitating subsequent heart rate calculation.The proposed method is validated on different datasets,and experimental results demonstrate that it significantly improves the accuracy of prediction and has a higher signal-to-noise ratio. |
| Key words: non-contact heart rate detection remote photoplethysmography generative adversarial network |
