Towards open-set myoelectric gesture recognition via dual-perspective inconsistency learning

Area of Science:

• Biomedical Engineering
• Human-Computer Interaction
• Machine Learning

Background:

• Surface electromyography (sEMG) is crucial for human-machine interaction (HMI), particularly in prosthetics and rehabilitation.
• Current sEMG gesture recognition struggles with open-set scenarios, failing to reject unknown gestures.
• Robust open-set recognition is vital for practical HMI applications.

Purpose of the Study:

• To develop an effective open-set gesture recognition system for sEMG.
• To address the challenge of distinguishing known gestures from unknown ones.
• To enhance the performance of myoelectric gesture recognition.

Main Methods:

• Discovered prediction inconsistency for unknown classes as a key factor for open-set recognition.
• Proposed PredIN, a dual-perspective inconsistency learning approach.
• Magnified prediction inconsistency by enhancing feature distribution differences across perspectives while optimizing intra-perspective separability.

Main Results:

• Evaluated PredIN on four public sEMG datasets.
• Demonstrated that PredIN significantly outperforms existing state-of-the-art methods.
• Achieved superior performance in open-set gesture recognition tasks.

Conclusions:

• The proposed PredIN method enables accurate closed-set gesture classification.
• PredIN effectively rejects unknown gestures, showcasing its superiority in open-set scenarios.
• This approach enhances the reliability and applicability of sEMG-based HMI.