ECNN-CS: Efficient Convolutional Neural Network by Channel and Spatial Fusion for Hand Gesture Recognition Using sEMG Sensors

Sensor-based gesture recognition technology has been widely applied in human-machine interaction, human-robot interaction, prosthetic control, medical rehabilitation, and other fields. Surface electromyography (sEMG) provides rich information about muscle motion that accurately reflects a user’s motion intention, making sEMG sensors critical for assistive and rehabilitative technologies. Therefore, sEMG-based gesture recognition has received extensive research attention. In this study, we propose a gesture recognition and classification method based on sEMG signals, referred to as ECNN-CS. The method inputs extracted features into an efficient convolutional neural network (ECNN). It incorporates both channel attention (CA) and spatial attention (SA) mechanisms to improve gesture classification at the decision layer. We evaluated ECNN-CS on the official DB4 and DB5 datasets. In the DB4 dataset (sampling frequency 2000Hz), we achieved an accuracy of up to 82.32% across 53 gesture recognition tasks, using a window length of 250ms (corresponding to 500 time steps) and a movement length of 62.5ms (125 time steps). In the DB5 dataset (sampling frequency 200Hz), the window length is 250ms (50 time steps), the movement length is 62.5ms (12 time steps), and the highest accuracy reaches 88.13%. These results demonstrate that ECNN-CS effectively leverages both temporal and spatial features of sEMG signals, achieving excellent performance in gesture recognition tasks and providing a solid foundation for future research and applications.