Dual Stream Long Short-Term Memory Feature Fusion Classifier for Surface Electromyography Gesture Recognition

Area of Science:

• Biomedical Engineering
• Human-Computer Interaction
• Machine Learning

Background:

• Electromyography (EMG) signal recognition is crucial for intelligent prosthetics and human-computer interaction.
• Current machine learning and deep learning methods face challenges like manual feature extraction, overfitting, and low adaptability.
• Existing deep learning models often use complex architectures, leading to computational inefficiency and potential accuracy limitations.

Purpose of the Study:

• To develop a novel, lightweight model for improved EMG-based hand gesture recognition.
• To enhance classification accuracy and reduce computational cost compared to existing methods.
• To enable more effective and efficient control of intelligent prosthetics through gesture recognition.

Main Methods:

• Proposed a dual stream LSTM feature fusion classifier integrating five time-domain EMG features and raw data.
• Employed one-dimensional convolutional neural networks (CNNs) and Long Short-Term Memory (LSTM) layers for feature processing and classification.
• Utilized a simple yet effective architecture to capture global EMG signal features with reduced computational demands.

Main Results:

• Achieved 89.66% accuracy on the public DB1 dataset (52 gestures, 27 subjects).
• Demonstrated a fast inference time of 87.6 ms per gesture, suitable for real-time applications.
• Validated on the DB2 dataset, achieving a subject-wise mean accuracy of 91.74%.

Conclusions:

• The proposed dual stream LSTM model effectively fuses time-domain features and raw EMG data for enhanced information extraction.
• The lightweight architecture provides an efficient and adaptable solution for EMG gesture recognition.
• The model's performance is comparable to complex deep learning networks, offering a practical approach for real-time prosthetic control.