A Multi-Scale CNN for Transfer Learning in sEMG-Based Hand Gesture Recognition for Prosthetic Devices.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Human-Computer Interaction

Background:

• Surface Electromyography (sEMG) is crucial for hand gesture recognition.
• Current deep learning models lack robustness and user adaptability.
• High computational resources are often required for sEMG models.

Purpose of the Study:

• To develop a robust and adaptive model for sEMG-based hand gesture recognition using Transfer Learning.
• To improve inter-subject generalization and enable quick adaptation to new users.
• To investigate Transfer Learning frameworks for user-dependent fine-tuning.

Main Methods:

• Proposed a Multi-Scale Convolutional Neural Network (MSCNN).
• Employed domain adaptation with a gradient-reversal layer and self-supervision using triplet margin loss for pre-training.
• Evaluated approaches on NinaPro benchmark datasets and compared Transfer Learning frameworks.
• Examined the impact of rectification and window length on performance.

Main Results:

• The proposed MSCNN with Transfer Learning demonstrated strong inter-subject generalization.
• One Transfer Learning framework achieved a 97% F1 Score across 14 classes with 1.40 epochs.
• Real-time accessible normalizing techniques significantly improved usability and performance.
• The study confirmed the effectiveness of Transfer Learning for adaptive, user-specific models.

Conclusions:

• Transfer Learning is highly effective for creating adaptive, user-specific models for sEMG-based prosthetic hands.
• The developed model shows potential for on-site retraining to address performance degradation.
• Optimized preprocessing techniques enhance the real-world applicability of sEMG gesture recognition.