Fusion of Multi-Stream Convolution and Bidirectional Gated Recurrent Network for Electromyographic Gesture Recognition

Gesture recognition based on surface electromyography has become increasingly important in human-machine interaction. Existing deep learning architectures primarily rely on convolutional neural networks (CNNs) to capture spatial in-formation from sEMG signals. Given the multi-channel nature of temporal sEMG signals, we propose a neural network architecture that combines the power of multi-stream convolution and bidirectional gated units to better capture the temporal properties of multi-channel sEMG signals for gesture recognition. Moreover, individual differences in surface electromyography datasets can lead to overfitting and other issues in classification models. Data augmentation proves effective in mitigating overfitting problems by increasing dataset diversity, enabling the model to learn more comprehensive features, enhancing its generalization capability, and reducing the risk of overfitting. Experimental results on the NinaProDB2 dataset demonstrate that the proposed method achieves an average recognition accuracy of 89.21% on sparse multi-channel sEMG databases, surpassing the recognition level of traditional machine learning models. Additionally, the conducted ablation experiments validate the correctness of the proposed model.