Lite-MSTFNet: A lightweight sEMG gesture recognition network BASED on multi-scale spatiotemporal fusion and dynamic gated attention

Surface electromyography (sEMG) has emerged as a promising modality for gesture recognition in deep learning applications. However, deploying state-of-the-art deep learning models on resource-constrained embedded devices remains a critical challenge due to their substantial computational demands. To address this, we propose Lite-MSTFNet, a lightweight multi-scale spatio-temporal fusion network specifically tailored for efficient sEMG-based gesture recognition on edge devices. The architecture leverages a parallel multi-branch design with depthwise separable convolutions to effectively capture multi-scale spatio-temporal features while minimizing computational overhead. Furthermore, we introduce an adaptive dynamic gating attention mechanism that intelligently routes computational paths based on input complexity, thereby enabling efficient resource allocation. To address the inherent challenges of subject variability and signal non-stationarity, we develop a compact hyperparameter search strategy with a constrained search space for subject-specific optimization. Experimental validation on the Ninapro DB2 dataset demonstrates that our approach achieves competitive performance: an average recognition accuracy of 85.42% with merely 4.3 K model parameters and best-subject accuracy of 92.47%, while the dynamic attention mechanism operates on only 2.29% of test samples at a gate threshold of 0.7, theoretically reducing FLOPs by 97.5%. This work presents a practical and efficient solution for sEMG-based gesture recognition in resource-constrained scenarios. With only 4.3 K parameters and 0.05M FLOPs, the model’s theoretical complexity is substantially lower than previously deployed sEMG models on ARM Cortex-M microcontrollers, suggesting strong potential for embedded deployment pending hardware validation.