A Convolutional Neural Network with Narrow Kernel and Dual-View Feature Fusion for sEMG-Based Gesture Recognition

For gesture recognition based on surface Electromyography (sEMG), Convolutional Neural Network (CNN)-based techniques have made prominent progress to gesture features extract for motor rehabilitation or bionic hand control. However, the performance of traditional CNN algorithms shows limited capability to characterize the time-varying properties and channel independence of sEMG. In this paper, a Narrow Kernel and Dual-view Feature Fusion Convolutional Neural Network (NKDFF-CNN) is proposed. The Narrow Kernel was firstly used to fully extract the temporal features of each channel to protect the channel independence of sEMG. Feature fusion between two views to effectively reduce the number of the model parameters and avoid the overfitting phenomenon. Then, shallow and deep features were fused in the decision phase to improve the multi-dimensional characterization of sEMG. Finally, a more powerful Arcface loss function replaced Softmax for gesture classification. The algorithm was validated experimentally on the Ninapro DB2 dataset using a 200 ms window length. The results showed that NKDFF-CNN achieved an average recognition accuracy of 87.96% for 40 subjects, which is 0.94% higher than the best result using other methods. This algorithm presented in this paper effectively improved the performance of gesture recognition based sEMG and provided a new idea for this research direction.