A normalisation approach improves the performance of inter-subject sEMG-based hand gesture recognition with a ConvNet

Area of Science:

• Biomedical Engineering
• Machine Learning
• Human-Computer Interaction

Background:

• Subject-specific surface electromyography (sEMG) based gesture classification using deep learning is common.
• Collecting subject-specific training data for sEMG gesture recognition is time-consuming and impractical.
• Utilizing data from multiple subjects can improve classifier performance but requires effective normalization.

Purpose of the Study:

• To develop a real-time, subject-independent sEMG-based hand gesture classification method.
• To address the challenge of acquiring extensive subject-specific training data.
• To enable the use of generalized sEMG data for accurate gesture recognition.

Main Methods:

• Proposed a min-max normalization approach for sEMG data.
• Applied normalization to source domain data using target domain (new user) trial cycle values, assigned by class label.
• Utilized a convolutional neural network (ConvNet) for classification.
• Employed leave-one-subject-out cross-validation for evaluation.

Main Results:

• Achieved an average accuracy of 87.03% on the G. dataset (12 gestures).
• Achieved an average accuracy of 94.53% on the M. dataset (7 gestures).
• Demonstrated the effectiveness of the subject-independent normalization technique.

Conclusions:

• The proposed normalization method allows for real-time, subject-independent sEMG-based hand gesture classification.
• This approach significantly reduces the need for subject-specific training data.
• The method shows high accuracy across different datasets, indicating its robustness.