基于改进ResNet50的表面肌电信号手势识别  

SEMG gesture recognition based on improved ResNet50

作　　者：牛群峰[1] 石磊 贾昆明 桂冉冉 董鹏豪 王莉[1] Niu Qunfeng;Shi Lei;Jia Kunming;Gui Ranran;Dong Penghao;Wang Li(College of Electrical Engineering,Henan University of Technology,Zhengzhou 450000,China)

机构地区：[1]河南工业大学电气工程学院,郑州450000

出　　处：《国外电子测量技术》2024年第4期181-189,共9页Foreign Electronic Measurement Technology

基　　金：河南工业大学创新基金计划(2022ZKCJ03);河南省科技研究计划(2013000210100)项目资助。

摘　　要：为了提高手势动作在类别众多且相似度高的情况下的识别精度,提出了一种基于连续小波变换和残差神经网络Res-Net50的表面肌电信号手势识别方法。首先对Ninapro DB2和DB3的原始表面肌电信号进行预处理和连续小波变换,得到Multi-sEMG Wavelet Map数据集,然后送入改进的ResNet50模型进行识别分类。实验结果表明,改进后的ResNet50网络模型在Multi-sEMG Wavelet Map DB2和DB3中17种手势动作的平均准确率分别达到了96.40%和94.11%,相比ResNet50网络模型方法提升了4.87%和5.83%。实现了手势动作在类别繁多、相似度较高的情况下的精准识别。为基于非侵入式传感器和机器学习控制的假肢手提供了新方案。To enhance the recognition accuracy of gesture actions in scenarios with numerous and highly similar categories,a gesture recognition method for surface electromyographic signals based on continuous wavelet transform and residual neural network ResNet50 is proposed.The raw surface EMG signals collected are first preprocessed and continuously wavelet transformed to obtain the Multi-sEMG Wavelet Map dataset,and then fed into the improved ResNet50 model for recognition and classification.The experimental results show that the improved ResNet50 network model achieves an average accuracy of 96.40%and 94.11%for 17 gesture actions in Ninapro DB2 and DB3,respectively,which is an improvement of 4.87%and 5.83%compared to the ResNet50 network model method.Achieved accurate recognition of gesture actions in situations with numerous and highly similar categories.A new scheme is provided for prosthetic hand based on non-invasive sensors and machine learning control.

关 键 词：表面肌电信号 连续小波变换 Multi-sEMG Wavelet Map ResNet50 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] TN911.7[自动化与计算机技术—控制科学与工程]