Multi-View Fusion Network-Based Gesture Recognition Using sEMG Data.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Rehabilitation Medicine

Background:

• Surface electromyography (sEMG) is crucial for non-invasive rehabilitation medicine and human action recognition.
• Sparse sEMG analysis and multi-view fusion lag behind high-density sEMG, lacking methods to enrich feature information and reduce channel-wise data loss.

Purpose of the Study:

• To propose a novel deep learning framework for enhancing sparse sEMG feature information in multi-view fusion.
• To effectively reduce information loss in the channel dimension during feature extraction.
• To improve the accuracy and reduce individual variability in sEMG-based action recognition.

Main Methods:

• Developed a novel Inception-MaxPooling-Squeeze-Excitation (IMSE) network module to minimize feature information loss.
• Utilized multiple feature encoders with multi-core parallel processing for sparse sEMG feature map enrichment.
• Employed Swin Transformer (SwT) as the classification backbone and compared feature fusion effects across different decision layers.

Main Results:

• The proposed multi-view fusion network achieved 93.96% average accuracy in gesture action classification using a 300ms time window on the NinaPro DB1 dataset.
• Individual action recognition rate variation was limited to less than 11.2%, demonstrating reduced individuality differences.
• Fusion at decision layers significantly improved network classification performance.

Conclusions:

• The proposed multi-view learning framework effectively augments channel feature information from sparse sEMG signals.
• This approach offers a valuable reference for non-dense biosignal pattern recognition, particularly in reducing inter-subject variability.
• The study highlights the potential of advanced deep learning techniques for enhancing sparse sEMG applications in rehabilitation and human-computer interaction.