CNN-MSTINet: a lightweight hybrid network with multi-scale temporal interaction for sparse sEMG-based gesture recognition

Abstract Accurate gesture recognition from sparse surface electromyography (sEMG) signals remains challenging due to limited electrode coverage, vulnerability to noise, inter-session variability, and the difficulty of simultaneously capturing multi-scale temporal dynamics and subtle inter-channel relationships, particularly when distinguishing anatomically similar gestures that produce nearly indistinguishable muscle activation patterns under sparse electrode configurations. This study aims to address these measurement-inherent limitations through a lightweight, measurement-aware architecture suitable for real-time deployment on resource-constrained edge devices. We present CNN-MSTINet, a lightweight hybrid model that embeds measurement-aware design principles into its architecture. It integrates convolutional and Transformer components with: (1) parallel multi-resolution temporal convolutions with feature reuse to capture cross-scale signal dynamics without relying on fixed spatial priors; (2) a hierarchically adaptive multi-scale positional encoding scheme that models both millisecond-scale transients and gesture-level temporal evolution; (3) a cross-branch interaction module enabling bidirectional refinement between local channel-specific features and global temporal context; and (4) end-to-end co-design with lightweight preprocessing to meet embedded latency and memory constraints. Evaluated on a custom sparse-channel SIA-16 dataset and the 49-class Ninapro DB2 benchmark, CNN-MSTINet achieves 95.62% and 85.77% accuracy, respectively. The model demonstrates robustness to realistic noise (maintaining performance down to 0 dB SNR), fast convergence, and strong cross-subject generalization. Ablation studies quantitatively validate the contribution of each component, while interpretability analyses link decoding errors, such as the confusion between gesture 7 and gesture 8, to intrinsic measurement ambiguities rather than model deficiencies. CNN-MSTINet advances sEMG-based gesture recognition by explicitly modeling the physical and physiological constraints of the measurement process. It delivers high accuracy, noise resilience, and computational efficiency (⩽25.2 ms latency on STM32F407, 0.84–1.28 M parameters), making it well suited for real-time wearable human–machine interfaces where trust, interpretability, and deployability are paramount.