Abstract
Hand gesture classification using surface electromyography (sEMG) is fundamental for prosthetic control and human-machine interaction. However, most existing studies focus on high-density recordings or large gesture sets, leaving limited evidence on performance in low-channel, reduced-gesture configurations. This study addresses this gap by comparing a classical convolutional neural network (CNN), inspired by Atzori's design, with a Convolutional Vision Transformer (CViT) tailored for compact sEMG systems. Two datasets were evaluated: a proprietary Myo-based collection (10 subjects, 8 channels, six gestures) and a subset of NinaPro DB3 (11 transradial amputees, 12 channels, same gestures). Both models were trained using standardized preprocessing, segmentation, and balanced windowing procedures. Results show that the CNN performs robustly on homogeneous signals (Myo: 94.2% accuracy) but exhibits increased variability in amputee recordings (NinaPro: 92.0%). In contrast, the CViT consistently matches or surpasses the CNN, reaching 96.6% accuracy on Myo and 94.2% on NinaPro. Statistical analyses confirm significant differences in the Myo dataset. The objective of this work is to determine whether hybrid CNN-ViT architectures provide superior robustness and generalization under low-channel sEMG conditions. Rather than proposing a new architecture, this study delivers the first systematic benchmark of CNN and CViT models across amputee and non-amputee subjects using short windows, heterogeneous signals, and identical protocols, highlighting their suitability for compact prosthetic-control systems.