Lightweight Gesture Recognition Based on Depthwise Separable Convolution and FECAM Attention Mechanism for sEMG

Abstract

Surface electromyography (sEMG) is a promising approach for noninvasive gesture recognition in human–computer interaction and rehabilitation. However, existing high-accuracy models often incur high-computational costs, thereby limiting real-time deployment. To address this, we propose FSGR-Net, a lightweight residual network that reconstructs ResNet50 using a small-convolution stacking strategy and a Lite-Fusion Block. The Lite-Fusion Block integrates depthwise separable convolution (DSC), ghost convolution (GC), and a channel compression–expansion mechanism to reduce redundancy. In particular, a frequency-enhanced channel attention mechanism (FECAM) is introduced after DSC layers to enhance discriminative features while mitigating the Gibbs phenomenon. Furthermore, a joint data augmentation strategy—time-shifting and masking—is applied to improve generalization. Evaluations on NinaPro DB1, DB5, and our SC-Myo datasets show that FSGR-Net achieves 93.17%, 87.83%, and 93.35% accuracy, respectively, with only 0.85 M parameters and 0.22 G FLOPs, demonstrating strong potential for deployment in mobile and low-power wearable systems.