sEMG-based gesture recognition using multi-stream adaptive CNNs with integrated residual modules

Area of Science:

• Biomedical Engineering
• Machine Learning
• Human-Computer Interaction

Background:

• Surface electromyography (sEMG) is crucial for gesture recognition, but current deep learning models struggle with effective feature extraction, especially for sparse signals and multi-gesture scenarios.
• Existing convolutional neural networks (CNNs) often exhibit limitations in capturing intricate patterns within sEMG data, impacting overall recognition performance.

Purpose of the Study:

• To develop an advanced deep learning model for enhanced sEMG gesture recognition.
• To overcome the challenges of insufficient feature extraction and low accuracy in recognizing sparse and multi-gestures from sEMG signals.

Main Methods:

• Proposed a multi-stream adaptive convolutional neural network with residual modules (MSACNN-RM).
• Integrated multiple CNN streams, adaptive convolutional layers, and residual modules to boost feature extraction and learning capabilities.
• Leveraged multi-stream convolution and adaptive modules combined with ResNet blocks to extract crucial gesture features from sparse sEMG signals.

Main Results:

• Achieved high recognition accuracies: 98.24% on Ninapro DB1, 93.52% on Ninapro DB2, and 92.27% on Ninapro DB4.
• Demonstrated superior performance compared to existing deep learning models in sEMG gesture recognition.
• Effectively improved the model's ability to extract and understand complex data patterns from sEMG signals.

Conclusions:

• The MSACNN-RM model shows significant promise for accurate and robust sEMG gesture recognition.
• Future work should focus on developing universal algorithms to address inter-individual variations in sEMG signals and optimizing the network for reduced computational load.