sEMG Gesture Recognition Network Based on Time-Frequency Domain with Dual-Branch Feature Fusion

Gesture recognition via Surface Electromyography (sEMG) is vital in prosthetic control, rehabilitation, and human-computer interaction. However, the complex spatiotemporal nature of sEMG limits existing methods in feature extraction, similar gesture discrimination and robustness in real-world applications. To address this, we propose GPF-Net, a dual-branch gesture recognition network based on time-frequency domain. Continuous wavelet transform extracts multiscale time–frequency representations to enhance similar gesture distinction. The proposed network employs a dual-branch structure to overcome single-branch limitations and effectively capture local and global features. The Group Temporal Convolutional Network branch (GTCN) captures local details across frequencies, while the Positional Encoding and Multi-head Attention branch (PEMA) models long-distance dependencies. To effectively integrate multi-scale features, a Multi-Scale Attention Fusion module (MSAF) dynamically adjusts feature weights through an attention mechanism to suppress noise and enhance discriminative information. Experiments on Ninapro DB5 and extended Similar Gesture Dataset (SGDataset) achieve 91.18% and 95.94% recognition accuracies respectively, validating the method’s effectiveness and robustness.