| 赵昶辰* **,曹鹏程*,黄斌***,冯远静*.融合部分-整体时间窗口及双注意力机制的视频心率提取网络[J].高技术通讯(中文),2026,36(8):789~802 |
| 融合部分-整体时间窗口及双注意力机制的视频心率提取网络 |
| Video-based heart rate extraction network via partial-global time window dual attention fusion |
| |
| DOI：10. 3772 / j. issn. 1002 - 0470. 2026. 08. 002 |
| 中文关键词: 远程光体积描记； 神经网络； 心率估计； 分离时间窗口； 视频处理 |
| 英文关键词: remote photoplethysmography, neural network, heart rate estimation, separated time window, video processing |
| 基金项目: |
| 作者 | 单位 | | 赵昶辰* ** | （* 浙江工业大学信息工程学院杭州 310023）
（**杭州电子科技大学计算机学院杭州 310018）
（***北京航空航天大学杭州创新研究院杭州 310051） | | 曹鹏程* | | | 黄斌*** | | | 冯远静* | |
|
| 摘要点击次数: 21 |
| 全文下载次数: 16 |
| 中文摘要: |
| 虽然近年来出现了许多远程光体积描记（remote photoplethysmography，rPPG）心率提取神经网络，但是大多只考虑了时空连续性而忽略了不可预测的视频数据异常情况，导致心率识别精准度下降。为此本文提出了一种基于部分-整体时间窗口特征融合与时间-通道联合注意力机制的视频心率提取网络，旨在当视频中出现异常情况时仍能够稳定、准确地估计心率，并将注意力机制与损失函数相结合定义联合损失函数。所提网络在PURE和UBFC-rPPG这2个数据集上与现有的9种方法进行比较，以验证所提模型的预测能力。实验结果表明，在PURE数据集上，相比于次优的模型，所提方法其信噪比（signal-to-noise ratio，SNR）提高了25.21%，平均绝对误差（mean absolute error，MAE）降低了31.63%，均方根误差(root mean square error，RMSE)降低了34.00%，ρ提高了4.25%；在UBFC-rPPG数据集上，相比于次优的模型，所提方法其SNR提高了13.62%，MAE降低了15.38%，RMSE降低了47.05%，ρ提高了0.40%。实验证明融合部分-整体时间窗口特征以及时间-通道联合注意力机制可有效抑制噪声干扰，进一步提升心率测量精度。 |
| 英文摘要: |
| While numerous remote photoplethysmography (rPPG) heart rate extraction neural networks have emerged in recent years, most of them only consider spatiotemporal continuity and neglect unpredictable anomalies in video data, leading to a decrease in heart rate measurement accuracy. This paper proposes a video-based heart rate extraction network based on the fusion of partial-global temporal window features and a time-channel joint attention mechanism. Its aim is to accurately estimate heart rate in the presence of abnormal situations in the video, and to define a hybrid loss function by combining the attention mechanism loss. The proposed network is compared with nine existing methods on PURE and UBFC-rPPG datasets to validate its performance. Experimental results show that on the PURE dataset, compared with the second-best model, signal-to-noise-ratio (SNR) of the proposed network is improved by 25.21%, mean absolute error (MAE) decreased by 31.63%, root mean square error (RMSE) decreased by 34.00%, and ρ increased by 4.25%. On the UBFC-rPPG dataset, compared to the second-best model, the SNR of the proposed network is improved by 13.62%, MAE decreased by 15.38%, RMSE decreased by 47.05%, and ρ increased by 0.40%. The experimental results demonstrate that the fusion of partial-global temporal window features and the time-channel joint attention mechanism can effectively suppress noise interference and further enhance heart rate measurement accuracy. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
