An action-level gesture recognition method based on sEMG and channel-recalibrated multiscale temporal features

Abstract

Gesture recognition based on surface electromyography (sEMG) signals is a key technology for wearable human-computer interaction and intelligent assistive systems. However, achieving robust gesture recognition remains challenging due to the limited temporal resolution associated with low sampling rates and the significant intra-gesture variability. To address these issues, this study proposes an action-level recognition framework named SE-MSCFormer. This framework integrates three core components: (1) multi-scale temporal convolutions designed to extract local activation patterns across varying receptive fields; (2) a "Squeeze-and-Excitation" mechanism that adaptively recalibrates feature channels to highlight critical muscle activation responses; and (3) a Transformer-based temporal module capable of capturing long-range dependencies within 200 Hz sEMG signals. Extensive evaluations were conducted on the publicly available NinaPro DB5 dataset (action groups E1–E3) using a rigorous "within-subject, repetition-based" experimental protocol. Across five experiments with different random seeds, SE-MSCFormer achieved an accuracy of 92.44% and a Macro-F1 score of 92.38% on 1,040 independent test gesture samples. These results outperform the baseline MSCFormer model (which lacks channel recalibration and achieved 90.38% accuracy) and the standard Transformer model (89.77% accuracy). Ablation studies and sensitivity analyses confirm that the synergy between the multi-scale temporal configuration and channel recalibration enhances overall recognition performance. The results demonstrate that the proposed SE-MSCFormer offers an efficient and robust solution for low-sampling-rate sEMG gesture recognition.