An efficient surface electromyography-based gesture recognition algorithm based on multiscale fusion convolution and...

Area of Science:

• Biomedical Engineering
• Rehabilitation Technology
• Machine Learning

Background:

• Deep learning models for surface electromyography (sEMG) based gesture recognition often suffer from high computational complexity, limiting their practical use in rehabilitation.
• Efficient and accurate gesture recognition is crucial for advancing assistive technologies and improving patient outcomes.

Purpose of the Study:

• To develop a computationally efficient and accurate deep learning model for multitype gesture recognition using sEMG signals.
• To address the limitations of existing models by reducing algorithmic complexity without sacrificing performance.

Main Methods:

• Proposed the Residual-Inception-Efficient (RIE) model, integrating Inception modules for multiscale feature extraction and efficient channel attention (ECA) for feature reweighting.
• Implemented fast dimensionality reduction, asymmetric convolution decomposition, and pooling within the Inception module to minimize parameters and complexity.
• Conducted experiments on NinaPro DB1, DB3, and DB4 datasets for 52, 49, and 52-class gesture recognition, respectively.

Main Results:

• The RIE model achieved high recognition accuracies: 88.27% on NinaPro DB1, 69.52% on DB3, and 84.55% on DB4.
• Demonstrated excellent generalization ability across different datasets and gesture complexities.
• Significantly reduced algorithmic complexity in both spatial and temporal dimensions, resulting in a smaller model size and faster computation compared to other lightweight algorithms.

Conclusions:

• The RIE model offers a practical solution for sEMG-based gesture recognition, balancing lightweight computational requirements with reliable performance.
• This efficient deep learning approach holds significant potential for real-world applications in rehabilitation and human-computer interaction.