EMG-FRNet: A feature reconstruction network for EMG irrelevant gesture recognition

Area of Science:

• Biomedical Engineering
• Machine Learning
• Human-Computer Interaction

Background:

• Surface electromyography (EMG) signals are crucial for gesture recognition in human-computer interaction.
• Current systems struggle with interference from irrelevant gestures, impacting accuracy and security.
• Detecting irrelevant gestures is vital for robust EMG-based systems.

Purpose of the Study:

• To develop an effective method for irrelevant gesture recognition using surface EMG signals.
• To introduce and adapt the GANomaly network for EMG-based irrelevant gesture detection.
• To propose an improved feature reconstruction network, EMG-FRNet, for enhanced performance.

Main Methods:

• Adapted the GANomaly network for anomaly detection in EMG signals, distinguishing target from irrelevant gestures based on feature reconstruction error.
• Proposed EMG-FRNet, incorporating channel cropping (CC), cross-layer encoding-decoding feature fusion (CLEDFF), and SE channel attention (SE).
• Validated the model on Ninapro DB1, Ninapro DB5, and a self-collected dataset.

Main Results:

• EMG-FRNet demonstrated high performance across all tested datasets.
• Achieved Area Under the receiver operating characteristic Curve (AUC) values of 0.940 (Ninapro DB1), 0.926 (Ninapro DB5), and 0.962 (self-collected).
• Outperformed existing related research in EMG irrelevant gesture recognition accuracy.

Conclusions:

• The proposed EMG-FRNet model significantly improves the accuracy and reliability of irrelevant gesture recognition in EMG-based systems.
• The integration of GANomaly principles with novel network structures offers a promising approach for enhancing human-computer interaction security.
• EMG-FRNet represents a substantial advancement in addressing interference from unwanted movements in EMG gesture recognition.