sEMG-Based Gesture Recognition Using Deep Learning From Noisy Labels.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Signal Processing

Background:

• Myoelectric prosthesis control using surface Electromyography (sEMG) faces challenges in accuracy and optimal performance.
• Current Muscle-Computer Interface (MCI) systems require improved input-output flexibility and dataset quality for better gesture recognition.

Purpose of the Study:

• To propose a novel gesture recognition framework to enhance the performance of MCI for myoelectric prosthesis control.
• To address the limitations of sparse sEMG signals and label noise in existing systems.

Main Methods:

• Developed a novel framework integrating multiresolution wavelet decomposition to create a fused sEMG-Map image, enriching spectral information.
• Employed a Convolutional Neural Network (CNN) to process the sEMG-Map, leveraging its ability to exploit hierarchical features.
• Introduced an Adaptive Label Refinement (ALR) strategy within a CNN (ALR-CNN) to simultaneously correct mislabeled data and optimize the model.

Main Results:

• Achieved high average classification accuracies: 95.50% on NinaPro DB2, 95.85% on NinaPro DB7, and 85.58% on NinaPro DB3.
• The proposed sEMG-Map and ALR-CNN effectively improved gesture recognition performance on large-scale public datasets.
• Experimental results validated the framework's efficacy and demonstrated significant accuracy improvements.

Conclusions:

• The novel framework effectively enhances spectral information from sparse sEMG signals and robustly handles label noise.
• The proposed method offers a significant advancement in myoelectric prosthesis control, achieving high accuracy and demonstrating practical utility.