sEMG-Based Gesture Recognition via Multi-Feature Fusion Network

The sparse surface electromyography-based gesture recognition suffers from the problems of feature information not richness and poor generalization to small sample data. Therefore, a multi-feature fusion network (MFF-Net) model is proposed in this paper. This network incorporates long short-term memory (LSTM) and the attention mechanism into the model, and three sub-networks are constructed for enhancement of features in the time, frequency and spatial domains, respectively. The introduced attention mechanism enhances useful features and weakens useless ones. Then, the processed features are spliced and stacked, which strengthens the information between time and channel to enrich features in sparse sEMG, improved model performance for feature processing. To further validate that the proposed model is effective in improving gesture recognition accuracy. We selected 18 gesture recognition tasks from the NinaPro DB3 and DB7 datasets for experimental evaluation. Among them, ablation experiments were conducted on intact subjects data in DB7. The experimental results show that the proposed model reaches the current optimal in gesture recognition, with 92.47% classification accuracy. Moreover, the model can be transferred to gesture recognition for small sample amputees data, which is also effective in solving insufficient data problem. Two amputees (in DB7) recognition rate have significantly improved from 60.35% to 84.93%, while eleven amputees (in DB3) recognition rate have significantly improved from 71.84% to 82.00%. It is demonstrated the applicability and generalization of the proposed model transfer learning to the amputees gesture recognition task.