Gesture Recognition for Prosthetic Hand Control Using Surface Electromyography Time-Frequency Images and Convolutional Neural Networks (Demonstration video and paper data)

Abstract: Surface electromyography (sEMG) signals, as a non-invasive biosignal, are widely used in human-machine interaction and prosthetic control systems. Gesture recognition using sEMG is one of the key technologies for electromyographic prosthetic hand control systems. This study proposes a gesture recognition method based on sEMG time-frequency images and convolutional neural networks (CNN), successfully applied to real-time control of various prosthetic hand gestures. sEMG data were collected and preprocessed from 12 healthy subjects, with short-time Fourier transform (STFT) used to generate time-frequency images. Five CNN models with different depths were constructed for sEMG gesture recognition. Experimental results showed that the 4-layer deep CNN model (CNN4) performed the best in 9 gesture classification tasks, achieving an accuracy of 96.88%. Among classic CNN architectures, most models achieved classification accuracy above 90%, with DenseNet performing the best at 96.81%. Additionally, the study investigated the impact of time-frequency image resolution and STFT window function time window length on CNN performance. The results demonstrated that appropriate time-frequency image resolution and window length significantly improved CNN recognition accuracy. This method successfully improves gesture recognition accuracy and real-time performance by optimizing the input of sEMG time-frequency images and CNN network structure, contributing to the advancement of myoelectric prosthetic hand control systems.