| 引用本文: | 肖城钢,闵华松.基于全局与局部肌电特征交互的手势识别网络[J].控制理论与应用,2025,42(3):609~617.[点击复制] |
| XIAO Cheng-gang,MIN Hua-song.Gesture recognition network based on the interaction of global and local myoelectric features[J].Control Theory and Technology,2025,42(3):609~617.[点击复制] |
|
| 基于全局与局部肌电特征交互的手势识别网络 |
| Gesture recognition network based on the interaction of global and local myoelectric features |
| 摘要点击 35 全文点击 2 投稿时间：2023-04-10 修订日期：2024-11-23 |
| 查看全文 查看/发表评论 下载PDF阅读器 |
| DOI编号 10.7641/CTA.2023.30204 |
| 2025,42(3):609-617 |
| 中文关键词 表面肌电信号 手势识别 空洞卷积网络 注意力机制 特征融合 |
| 英文关键词 surface electromyography hand gesture recognition dilated convolutional networks attention mechanism feature fusion |
| 基金项目 国家自然科学基金项目(62073249), 国家重点研发计划项目(2022YFB4700400)资助. |
|
| 中文摘要 |
| 为了更有效地捕捉肌电信号中的长期动态依赖关系和局部细节信息,减少固有肌电特征信息损失对手势分类精度的影响,本文提出一种基于全局–局部特征交互的手势识别网络GL-EMG-Net.首先,融合空洞卷积和多头自注意力机制,设计全局特征提取模块Global-DT,提取肌电信号中的全局信息;然后,借助深度可分离卷积和注意力机制,设计局部特征提取模块Local-SK捕捉肌电信号中不同尺度的局部细节信息,并将提取的细节信息通过反馈机制反馈至Global-DT模块,完成局部特征与全局特征的交互;最后,将全局特征与局部特征融合后进行分类.实验结果表明,该手势识别网络在Ninapro DB5数据集的52种手势和实际12种常见手势分类中,表现出较高的手势分类精度和较强的鲁棒性. |
| 英文摘要 |
| In order to capture the long-term dynamic dependencies and local detail information in the electromyography (EMG) signal more effectively and reduce the impact of the loss of intrinsic EMG feature information on the gesture classification accuracy, we propose a gesture recognition network the global and local-electromyography-network (GLEMG-Net) based on the interaction of global and local features. Firstly, the dilated convolution and multi-head selfattention mechanism are integrated to design the global feature extraction block the global-dilation transformer (Global-DT) to extract the global information in the EMG signal. Then, with the help of the depth separable convolution and attention mechanism, the local feature extraction block the local-selective kernel (Local-SK) is designed to capture the local detail information of different scales in the EMG signal, and feedback the extracted detail information to the Global-DT module through the feedback mechanism to complete the interaction between local features and global features. Finally, the global features and local features are fused for classification. The experimental results show that the gesture recognition network shows high gesture classification accuracy and strong robustness in the 52 gestures of Ninapro DB5 dataset and 12 actual common gestures. |
|
|
|
|
|