Surface-Electromyography-Based Gesture Recognition by Multi-View Deep Learning.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Human-Computer Interaction

Background:

• Surface electromyography (sEMG) based gesture recognition is crucial for muscle-computer interfaces but faces challenges with sparse, multichannel data.
• Existing methods often yield suboptimal performance, limiting the effectiveness of human-computer interaction.

Purpose of the Study:

• To develop an advanced multi-view deep learning framework for enhanced sEMG gesture recognition.
• To improve the accuracy and robustness of muscle-computer interfaces using sEMG data.

Main Methods:

• A novel multi-view convolutional neural network (CNN) framework combining classical sEMG features with deep learning.
• Utilizing a multistream CNN for parallel modeling of sEMG multi-view representations.
• Implementing a performance-based view construction strategy and a view aggregation network for feature fusion (early and late).

Main Results:

• The proposed multi-view framework significantly outperforms single-view methods in sEMG gesture recognition.
• Demonstrated superior performance on 11 sparse multichannel sEMG databases.
• Validated effectiveness on multimodal databases including sEMG and inertial measurement unit (IMU) data.

Conclusions:

• The multi-view deep learning approach offers a significant advancement for sEMG-based gesture recognition.
• This framework provides a more optimal solution for muscle-computer interface applications.
• Effective fusion of multi-view deep features enhances recognition accuracy in both unimodal and multimodal settings.