Improved Network and Training Scheme for Cross-Trial Surface Electromyography (sEMG)-Based Gesture Recognition.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Human-Computer Interaction

Background:

• Surface electromyography (sEMG) is crucial for non-invasive gesture recognition.
• Current sEMG methods struggle with trial-to-trial variability, limiting performance.
• Developing robust and accurate sEMG-based gesture recognition remains a significant challenge.

Purpose of the Study:

• To introduce a novel network-agnostic two-stage training scheme, sEMGPoseMIM, for enhanced sEMG gesture recognition.
• To improve the robustness of sEMG representations against trial variations.
• To align sEMG signals with corresponding hand movements using cross-modal knowledge distillation.

Main Methods:

• Proposed sEMGPoseMIM, a two-stage training scheme utilizing cross-trial mutual information maximization and knowledge distillation.
• Introduced sEMGXCM, a novel network designed as the sEMG encoder within the training scheme.
• Conducted experiments on seven diverse sparse multichannel sEMG datasets.

Main Results:

• The sEMGPoseMIM scheme achieved an average performance improvement of +1.3% on sparse multichannel sEMG databases compared to existing methods.
• The proposed sEMGXCM network, when trained from scratch, outperformed other networks by an average of +1.5%.
• Demonstrated the effectiveness of the proposed training scheme and network in producing trial-invariant sEMG representations.

Conclusions:

• The sEMGPoseMIM training scheme significantly enhances sEMG-based gesture recognition performance.
• The sEMGXCM network serves as an effective encoder for robust sEMG signal processing.
• The proposed methods offer a promising advancement for accurate and reliable human-computer interaction via sEMG.