A novel sEMG-based dynamic hand gesture recognition approach via residual attention network

With the emergence of more and more lightweight, convenient and cheap surface electromyography signal (sEMG) snsors, gesture recognition based on sEMG sensors has attracted much attention of researchers. In this study, combined with the sEMG sensor, a novel dynamic hand gesture recognition approach is proposed for effective and accurate dynamic gesture prediction. Here, a portable sEMG sensor (Myo wristband) is adopted to acquire the multi-channel sEMG signals of dynamic hand gestures and the continuous wavelet transformation (CWT) is proposed for data preprocessing to acquire the time-frequency maps. Due to the success of powerful contextual feature representation capability of deep convolutional neural networks (DCNNs), a deep residual attention network is proposed for accurate prediction of time-frequency maps. To effectively extract the key spatial and channel features from multi-channel sEMG signals, a residual attention network is proposed to act as the backbone network for effective feature representation. Besides, In the proposed recognition network, a multi-scale feature enhancement (MFE) module and an attention fusion block (AFB) are proposed, which respectively improve the multi-scale expression ability of the network and effectively realize multi-scale feature enhancement, respectively. Experimental results show that the proposed recognition network could achieve a superior detection ability compared with other state-of-the-art recognition models. The source code and dataset are available at https://github.com/lyangucas92/Ges_Net