Inverted-Xception: A novel network with multi-channel squeeze-and-excitation temporal deep convolution for high-precision surface electromyography-based gesture recognition

Gestures are vital in human-computer interaction, which have demonstrated substantial application potential in domains such as bionic prosthetic control and medical rehabilitation. Despite significant advancements in Surface Electromyography (sEMG) pattern recognition driven by deep learning, contemporary methodologies and models often exhibit limitations in capturing cross-channel dependencies which are crucial for recognizing the strong correlation between gesture actions and channel intensity distributions. To address these challenges, an innovative and transferable architecture named Inverted-Xception was proposed to integrate the Xception parallel framework with Inverted Residual Network (Inverted ResNet) by Multi-stream Fusion (MSF). The MSE-TDC module, specifically designed for sEMG, significantly enhances global feature representation and reduces computational cost while optimizing cross-channel attention allocation and spatial principal pattern perception. On the public NinaPro DB5 dataset, the proposed Inverted-Xception network achieved an average accuracy of 95.10 % across multiple subjects, outperforming previous models. Furthermore, with the improved transfer framework, Inverted-Xception overcomes individual variability and achieves cross-subject prediction performance ranging from 81.25 % to 92.19 %. Its outstanding performance establishes Inverted-Xception as a state-of-the-art solution for sEMG-based gesture recognition, offering promising prospects for future applications.