FS-HGR: Few-Shot Learning for Hand Gesture Recognition via Electromyography

Area of Science:

• Biomedical Engineering
• Machine Learning
• Human-Computer Interaction

Background:

• Deep Neural Networks (DNNs) excel in human-machine interfaces but require extensive data.
• Limited training data significantly degrades DNN performance for gesture detection.
• Collecting large datasets for real-world applications is often impractical.

Purpose of the Study:

• To design a DNN-based gesture detection model that achieves high accuracy with minimal training data.
• To address the challenge of data scarcity in training gesture recognition systems.
• To develop a practical solution for real-life human-machine interface applications.

Main Methods:

• Proposed a novel Few-Shot learning- Hand Gesture Recognition (FS-HGR) architecture.
• Combined temporal convolutions with attention mechanisms for effective generalization.
• Leveraged few-shot learning principles to infer outputs from few training observations.

Main Results:

• Achieved 85.94% accuracy on new repetitions (5-way 5-shot) on the Ninapro DB2 dataset.
• Reached 81.29% accuracy on new subjects (5-way 5-shot) for Ninapro DB2.
• Obtained 64.65% accuracy on new subjects (5-way 5-shot) for the Ninapro DB5 dataset.

Conclusions:

• The FS-HGR model demonstrates high accuracy in hand gesture recognition with limited data.
• The proposed architecture effectively generalizes from few training examples.
• This approach offers a viable solution for data-efficient gesture detection in real-world scenarios.