FeuNet: cross-user gesture recognition and fingertip force estimation network based on feature enhancement and U-Net

Area of Science:

• Biomedical Engineering
• Machine Learning
• Human-Computer Interaction

Background:

• Surface electromyography (sEMG) is crucial for gesture recognition and force estimation.
• Cross-user generalization in sEMG-based systems presents significant challenges.
• Existing methods often require extensive user-specific calibration data.

Purpose of the Study:

• To develop an advanced deep learning network, FeuNet, for improved cross-user sEMG gesture recognition and fingertip force estimation.
• To enhance sEMG representation and shallow feature mapping using a novel network architecture.
• To introduce an adaptive loss function for optimizing both recognition and estimation tasks simultaneously.

Main Methods:

• Proposed FeuNet, a deep learning network incorporating a feature enhancement module and a sEU-Net architecture.
• Implemented an adaptive loss function to balance gesture recognition and force estimation.
• Evaluated performance using 50% of new user data for calibration.

Main Results:

• FeuNet achieved 78.97% gesture recognition accuracy with 50% user calibration.
• Normalized root mean square error for force estimation was 5.51%.
• Coefficient of determination reached 67.12%, outperforming existing methods in cross-user scenarios.

Conclusions:

• FeuNet demonstrates superior cross-user generalization capabilities for sEMG-based applications.
• The proposed method significantly improves calibration efficiency by requiring less user data.
• FeuNet offers a promising solution for robust and efficient human-computer interaction using sEMG signals.