sEMG-based gesture recognition using multi-domain feature fusion with a lightweight Vanilla network

Gesture recognition based on surface electromyography (sEMG) has emerged as a promising approach for human–machine interaction systems, particularly in applications such as prosthetic hand control. Nevertheless, achieving an optimal balance between computational complexity and classification accuracy remains a persistent challenge for recognition networks. Thus, this paper proposes a multi-domain feature fusion (MDFF) methodology coupled with a lightweight Vanilla network (LVNet) to reduce computational demands whilst maintaining satisfactory classification performance, thereby enabling its direct deployment on terminal devices with limited computing resources, such as laptops or intelligent prosthetic hands. The proposed MDFF-LVNet model establishes an end-to-end fully convolutional classification architecture: the MDFF extracts and fuses time-domain and time–frequency domain features, and the LVNet improves inference speed and recognition capability by using dynamic convolution and activation functions to train dual convolutional operations. Experimental results demonstrate that the MDFF-LVNet model achieves classification accuracies of 95.78%, 92.77%, floating-point operations of 4.52 GFLOPs and 8.05 GFLOPs, and inference time of 18.11 ms and 45.69 ms on public gesture datasets NinaPro DB2 and DB5, respectively. To evaluate its online recognition performance, experiments conducted on a bionic prosthetic hand using a self-constructed sEMG dataset of 6 gestures achieved an offline accuracy of 99.57%.