Abstract
High-density surface electromyography (HD-sEMG) is a promising noninvasive modality for high-resolution gesture recognition. While hand-crafted features remain common, end-to-end deep learning has emerged as an effective approach to learn representations directly from raw signals. In this article, we present HGR-STAR, an end-to-end framework that adapts the StarNet to HD-sEMG analysis. We propose a novel HD-sEMG embedding module based on the short-time Fourier transform (STFT), which converts raw spatiotemporal signals into structured time-frequency representations while preserving electrodewise information. To improve generalization under limited data, label smoothing is incorporated for better confidence calibration. Evaluated on a public HD-sEMG dataset with 65 gesture classes, HGR-STAR achieves 95.21% average validation accuracy, outperforming or matching existing state-of-the-art methods, including long short-term memory (LSTM), convolutional neural network (CNN), graph neural network (GNN), and Vision Transformer variants. Ablation studies show that the full model improves accuracy by 0.57% and reduces performance variance (std: 5.01%, ↓ 0.52%) compared to a version without the STFT embedding and label smoothing. With a validation latency of 5 ms per batch (32 samples), HGR-STAR is well suited for real-time wearable applications.
| Original language | English |
|---|---|
| Journal | IEEE Transactions on Human-Machine Systems |
| Early online date | 9 Jul 2026 |
| DOIs | |
| Publication status | Early online - 9 Jul 2026 |
Keywords
- End-to-end learning
- hand gesture recognition
- high-density surface electromyography (HD-sEMG) embedding
- interdisciplinary modeling
- label smoothing
- StarNet
Fingerprint
Dive into the research topics of 'HGR-STAR: Raw HD-sEMG Gesture Recognition via StarNet With HD-sEMG Embedding and Label Smoothing'. Together they form a unique fingerprint.Cite this
- APA
- Author
- BIBTEX
- Harvard
- Standard
- RIS
- Vancouver