Knowledge distillation for sEMG-based gesture recognition: Enhancing wearable HMI systems with lightweight models

Area of Science:

• Biomedical Engineering
• Machine Learning
• Wearable Technology

Background:

• Surface electromyography (sEMG) is key for wearable human-machine interaction (HMI).
• Cross-user gesture recognition models struggle with individual physiological signal variability.
• Deep learning models offer generalization but are computationally demanding for wearables.

Purpose of the Study:

• To address the challenge of deploying deep learning models on resource-constrained wearable devices for sEMG gesture recognition.
• To investigate knowledge distillation as a method for compressing large gesture recognition models into lightweight versions.
• To evaluate the performance of distilled models in cross-user scenarios.

Main Methods:

• Trained deep teacher models (DenseNet, InceptionV3, VggNet).
• Distilled parameters from teacher models into compact Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) student models.
• Conducted experiments using sEMG data from 30 subjects collected via a 32-channel wristband electrode.

Main Results:

• Distilled student models, especially LSTM-based ones, achieved classification accuracy comparable to or exceeding teacher models.
• The DenseNet121-LSTM architecture demonstrated the highest classification accuracy among tested teacher-student combinations.
• Analysis confirmed that distilled models effectively approximate the performance of high-complexity (FLOPs) models, balancing performance and computational cost.

Conclusions:

• This research demonstrates the feasibility of implementing advanced deep learning for gesture recognition in wearable systems.
• Knowledge distillation enables efficient sEMG-based gesture recognition on devices like wristbands and smartwatches.
• The findings facilitate more responsive and practical daily-life applications leveraging wearable HMI.