Masked Autoencoder with Swin Transformer Network for Mitigating Electrode Shift in HD-EMG-based Gesture Recognition

TL;DR

This paper introduces MAST, a novel self-supervised learning framework using masked autoencoders and Swin Transformer architecture to enhance gesture recognition robustness against electrode shift in high-density sEMG signals.

Contribution

The paper proposes a new MAST framework that employs multiple masking strategies and a multi-path Swin-Unet architecture for improved electrode shift robustness in HD-sEMG gesture recognition.

Findings

MAST outperforms existing methods in robustness against electrode shift.

Multi-mask strategies enhance the learned representations.

Self-supervised pre-training improves generalization across sessions.

Abstract

Multi-channel surface Electromyography (sEMG), also referred to as high-density sEMG (HD-sEMG), plays a crucial role in improving gesture recognition performance for myoelectric control. Pattern recognition models developed based on HD-sEMG, however, are vulnerable to changing recording conditions (e.g., signal variability due to electrode shift). This has resulted in significant degradation in performance across subjects, and sessions. In this context, the paper proposes the Masked Autoencoder with Swin Transformer (MAST) framework, where training is performed on a masked subset of HDsEMG channels. A combination of four masking strategies, i.e., random block masking; temporal masking; sensor-wise random masking, and; multi-scale masking, is used to learn latent representations and increase robustness against electrode shift. The masked data is then passed through MAST's three-path encoder-decoder structure, leveraging a multi-path Swin-Unet architecture that simultaneously captures time-domain, frequency-domain, and magnitude-based features of the underlying HD-sEMG signal. These augmented inputs are then used in a self-supervised pre-training fashion to improve the model's generalization capabilities. Experimental results demonstrate the superior performance of the proposed MAST framework in comparison to its counterparts.