MSDS-FusionNet: A Multi-Scale Dual-Stream Fusion Network for High-Accuracy sEMG-Based Gesture Classification

Recently, deep-learning-based approaches have been widely applied in sEMG-based gesture recognition, however, existing methods typically focus on either time-domain or frequency-domain features, failing to leverage the complementary information from both domains, which limits the accuracy and robustness of gesture recognition systems. To address these limitations, we propose the Multi-Scale Dual-Stream Fusion Network (MSDS-FusionNet), a novel deep learning framework that integrates temporal and frequency-domain features for enhanced gesture classification accuracy. MSDS-FusionNet introduces two key innovations: the Multi-Scale Mamba (MSM) modules, which extract multi-scale temporal features through parallel convolutions with varying kernel sizes and linear-time sequence modeling with selective state spaces, enabling the capture of temporal patterns at multiple scales and the modeling of both short-term and long-term dependencies; and the Bi-directional Attention Fusion Module (BAFM), which effectively combines temporal and frequency-domain features using bi-directional attention mechanisms to fuse complementary information and improve recognition accuracy dynamically. Extensive experiments on the NinaPro dataset demonstrate that MSDS-FusionNet outperforms state-of-the-art methods, achieving accuracy improvements of up to 2.41%, 2.46%, and 1.38% on the DB2, DB3, and DB4 datasets, respectively, with final accuracies of 90.15%, 72.32%, and 87.10%. This study presents a robust and flexible solution for sEMG-based gesture recognition, effectively addressing the complexities of recognizing intricate gestures and offering significant potential for applications in prosthetics, virtual reality, and assistive technologies. The code of this study is available at https://github.com/hdy6438/MSDS-FusionNet.