Application Research on Optimization Algorithm of sEMG Gesture Recognition Based on Light CNN+LSTM Model

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Human-Computer Interaction

Background:

• Deep learning for gesture recognition using surface electromyography (sEMG) is crucial for advanced human-computer interaction.
• Existing models often require significant storage space, limiting their application on embedded systems.
• Multistate muscle action recognition demands high accuracy and efficient models.

Purpose of the Study:

• To develop and optimize a feature model for multichannel sEMG signals for accurate multistate muscle action recognition.
• To create a compact deep learning model suitable for embedded chip deployment with limited storage.
• To enhance the accuracy and applicability of gesture recognition in human-computer interaction.

Main Methods:

• Constructed a feature model using multidimensional sequential sEMG images.
• Combined Convolutional Neural Network (CNN) and Long-Term Memory (LTM) network for sEMG signal recognition.
• Utilized Fast Fourier Transform (FFT) and Root Mean Square (RMS) for feature data processing.

Main Results:

• Achieved a high recognition rate for sEMG signals processed with FFT and RMS.
• Demonstrated a 91.40% recognition accuracy for complex gestures.
• The optimized model has a compact size of 1MB, suitable for embedded applications.

Conclusions:

• The proposed feature model and optimization method enable accurate multistate sEMG signal recognition.
• The compact and high-precision deep learning model can effectively control artificial hands.
• This approach advances the development of efficient and accurate gesture recognition systems for embedded human-computer interaction.