Abstract
Recently, surface electromyogram (sEMG) has a trend with an increasing number of electrodes to compose a 2-dimension (D) electrode array, which is called high density sEMG (HD-sEMG). However, gesture recognition algorithm with HD-sEMG is still a challenge especially in real time recognition application. This paper researched several spatial attention modules and embedded them to the input layer of neural network. In this way, we can re-weight the input channel to get a better accuracy, robustness and interpretability. By utilizing the Group Convolution Neural Network (CNN), the gesture classification accuracy is improved by 4.44% and 2.71% in CapgMyo and CSL-HDEMG dataset respectively. This method is so efficient that it achieves only with 128 parameters, barely increasing the computational overhead. Meanwhile, we compared the performance in 1-D, 2-D and 3-D CNN, and found that our 1-D group CNN has great advantages in total computational overhead without the loss of accuracy. It provides a practical solution for real time gesture recognition application.
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call