Related Experiment Video
Updated: Jul 1, 2025
08:15
Capturing Dynamic Finger Gesturing with High-resolution Surface Electromyography and Computer Vision
Published on: March 28, 2025
480
Integrated block-wise neural network with auto-learning search framework for finger gesture recognition using sEMG
Shurun Wang1, Hao Tang2, Feng Chen3
1School of Electrical Engineering and Automation, Hefei University of Technology, Hefei, 230009, China; Graduate School of Medicine, Juntendo University, Tokyo, 1138421, Japan.
Artificial Intelligence in Medicine
|March 10, 2024
Summary
This study introduces an auto-learning search framework (ALSF) that generates optimal neural networks for surface electromyography (sEMG) based gesture recognition. The framework creates efficient models with state-of-the-art performance and reduced resource needs for muscle-computer interfaces.
Area of Science:
- Biomedical Engineering
- Machine Learning
- Neuroscience
Background:
- Accurate finger gesture recognition using surface electromyography (sEMG) is crucial for muscle-computer interfaces but remains challenging.
- High-performance deep learning models exist, but optimizing their architecture requires significant expertise and effort, hindering widespread adoption.
Purpose of the Study:
- To develop an automated approach for designing efficient and high-performing neural network architectures for sEMG-based gesture recognition.
- To overcome the limitations of manual network architecture tuning in deep learning for muscle-computer interfaces.
Main Methods:
- An auto-learning search framework (ALSF) was developed using reinforcement learning to automatically generate integrated block-wised neural networks (IBWNNs).
- The IBWNN architecture incorporates multi-branch convolutional blocks and triplet attention sub-blocks for feature extraction, alongside dimensional reduction layers.
Main Results:
- The ALSF-generated models achieved state-of-the-art performance on the Ninapro DB5 dataset, outperforming popular contemporary networks.
- The generated models demonstrated a reduced number of parameters, indicating lower resource consumption for practical deployment.
Conclusions:
- The proposed ALSF effectively automates the design of specialized neural networks for sEMG gesture recognition.
- The generated IBWNNs offer a promising solution for efficient and accurate muscle-computer interfaces with practical deployment advantages.
