Fourier-Wavelet convolutional network improving the accuracy of sEMG-based gesture recognition

Non-invasive surface electromyography (sEMG) signals can help recognize hand gestures, indicating broad potential for prosthetic control, wearable devices, medical rehabilitation, and human-computer interaction. Effective and efficient modeling is urgently needed to capture global and localized transient patterns of sEMG signals. This work introduces a Fourier-Wavelet Convolutional Network (FWCNet), a novel framework for high-accuracy sEMG-based gesture recognition. FWCNet extracts global spectral and localized time-frequency features according to a Fourier-based convolutional network (FCNet) and a wavelet-based convolutional network (WCNet), respectively. An adaptive exponential weighted fusion (AEWF) module integrates these branch-specific representations, thereby improving recognition performance. Experimental results show that FWCNet achieves classification accuracies of 0.9457 on the self-constructed dataset, 0.9890 on BioPatRec DB1, 0.9934 on BioPatRec DB2, and 0.8663 on BioPatRec DB3. Comparative experiments demonstrate the competitive performance of FWCNet, highlighting its potential applications in prosthetic control and medical rehabilitation.