Related Experiment Video
Updated: Mar 28, 2026
Capturing Dynamic Finger Gesturing with High-resolution Surface Electromyography and Computer Vision
Published on: March 28, 2025
Adversarially regularized transformer with channel-wise noise for robust hand gesture recognition using surface
Doheun Cha1, Dong-Gyu Lee2, Sangtae Ahn1
1School of Electronic and Electrical Engineering, Kyungpook National University, Daegu, 41566 South Korea.
Abstract:
Surface electromyography (sEMG) is a non-invasive method that captures electrical signals through electrodes placed on the skin surface, allowing hand gesture recognition by interpreting muscle activity patterns. Despite its advantages, sEMG signals face challenges from non-stationarity and environmental noise, which complicate reliable classification. To overcome these challenges, we propose a novel network that integrates adversarial training with noise perturbation, improving both robustness and generalization in sEMG-based hand gesture recognition. Our approach achieves state-of-the-art results on benchmark datasets consisting of NinaPro DB1, DB2, and DB4, testing practical applications in prosthetic control and human-computer interaction. In aggregated evaluations, we achieve performance improvements of 3.52%p (DB1), 1.47%p (DB2), and 5.47%p (DB4) compared to EMGHandNet. In subject-wise evaluations, we observe performance gains of 1.73%p (DB1) and 5.15%p (DB4), while maintaining stable performance on DB2 despite its larger and more variable subject pool. Adversarial training also markedly enhances robustness against attacks: under FGSM and PGD attacks, in subject-wise evaluations, performance gains of 1.73%p (DB1) and 5.15%p (DB4) were observed, while maintaining stability in DB2 despite its larger subject pool and outliers. Adversarial training significantly improved robustness against attacks: under fast gradient sign method (FGSM) and projected gradient descent (PGD) attacks ([Formula: see text]=0.01), our model with adversarial training retained 32.92% (FGSM) and 23.88% (PGD) accuracy in DB1, 12.05% (FGSM) and 5.97% (PGD) in DB2, and 18.06% (FGSM) and 12.41% (PGD) in DB4, consistently outperforming models without adversarial training. These results demonstrate that our approach enhances both classification accuracy and adversarial robustness, addressing critical limitations in real-world applications such as prosthetic controls.
