ICA-CNN: Gesture Recognition Using CNN With Improved Channel Attention Mechanism and Multimodal Signals
Shu Shen, Xuebin Wang, Mengshi Wu, Kang Gu, Xinrong Chen, Xinyu Geng
Abstract
As a promising technology in human–computer interface (HCI), deep learning has good application prospects for gesture recognition from surface electromyography (sEMG) signals. However, mainstream network structures with satisfactory recognition accuracy are too complex to be deployed on edge devices. The lightweight neural network tends to have difficulty in fulfilling the requirement of recognition accuracy with limited computing power. In order to make the neural network adapt to different applications as much as possible, we propose a flexible and modular method based on sEMG and acceleration signals for gesture recognition. Moreover, a lightweight convolutional neural network with the improved channel attention mechanism (ICA-CNN) can achieve satisfactory performance in recognition accuracy or inference speed for different applications. The experimental results show that the recognition accuracy of the proposed method on 49 gestures is 94.24% using sEMG and acceleration signals, much higher than the accuracy of the method only using sEMG signals. In addition, the single inference of the proposed method takes only 38.6 ms on the CPU.