A Lightweight Frozen Multi-Convolution Dual-Branch Network for Efficient sEMG-Based Gesture Recognition
Year: 2026
Type: article
Abstract: Gesture recognition is important for rehabilitation assistance and intelligent prosthetic control. However, surface electromyography (sEMG) signals exhibit strong non-stationarity, and conventional deep-learning models require long training time and high computational cost, limiting their use on resource-constrained devices. This study proposes a Frozen Multi-Convolution Dual-Branch Network (FMC-DBNet) to address these challenges. The model employs randomly initialized and fixed convolutional kernels for training-free multi-scale feature extraction, substantially reducing computational overhead. A dual-branch architecture is adopted to capture complementary temporal and physiological patterns from raw sEMG signals and intrinsic mode functions (IMFs) obtained through variational mode decomposition (VMD). In addition, positive-proportion (PPV) and global-average-pooling (GAP) statistics enhance lightweight multi-resolution representation. Experiments on the Ninapro DB1 dataset show that FMC-DBNet achieves an average accuracy of 96.4% ± 1.9% across 27 subjects and reduces training time by approximately 90% compared with a conventional trainable CNN baseline. These results demonstrate that frozen random-convolution structures provide an efficient and robust alternative to fully trained deep networks, offering a promising solution for low-power and computationally efficient sEMG gesture recognition.
Source: Sensors
DOI: 10.3390/s26020580
Institution: East China University of Technology
Language: English
Subfield: Biomedical Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $3,137
Award: G1277061969
OpenAlex ID: W7124298993