表面肌电信号的多流卷积网络融合手势识别方法

GESTURE RECOGNITION BASED ON MULTI-STREAM CONVOLUTION NETWORK FOR SURFACE EMG SIGNAL

为了提高基于表面肌电信号的手势动作识别的准确率,提出一种多流卷积融合的深度学习方法,通过对手臂肌肉区域间所产生的肌电信号图像进行多流表征,将表征所产生的多个肌电信号的子图像分别输入到构建的多流卷积网络中,由多流卷积网络分支对这些不同的肌电子图像进行特征提取和建模;通过融合网络层进行特征融合,将融合后的特征输入Softmax层进行手势动作的分类,输出得到概率最大的动作类型的标签,从而达到提升手势识别准确率的效果。实验结果表明用多流卷积的方法处理肌电信号的手势识别准确率比传统机器学方法高出17百分点。

In order to improve the accuracy of gesture recognition based on surface EMG signals,we propose a multi-stream convolution fusion deep learning.We performed the multi-stream characterization of the EMG signals generated between different muscle regions of the arm.The sub-images of multiple EMG signals were input into the constructed multi-stream convolutional network.The multi-stream convolutional network branches performed feature extraction and modeling of those different sub-images.Through the feature fusion in the fusion network layer,we input the fused features into the Softmax layer for the classification of gesture actions,and output the label of the action type with the greatest probability,so as to improve the accuracy of gesture recognition.The experimental result shows that the gesture recognition accuracy of EMG signal processed by multi stream convolution method is 17 percentage points higher than that of traditional machine methods.