EdgeNetTF for lightweight time frequency representation learning in multiclass sEMG hand gesture recognition

Abstract Surface electromyography (sEMG)-based hand gesture recognition plays a critical role in human-machine interaction, rehabilitation, and myoelectric control. However, many deep learning models improve recognition performance by increasing architectural complexity, which limits their suitability for wearable and resource-constrained applications. This study proposes EdgeNetTF, a lightweight time-frequency representation-learning model for multi-class sEMG hand-gesture recognition. EdgeNetTF employs two compact one-dimensional convolutional branches to learn complementary temporal and fast Fourier transform (FFT)-based spectral representations, followed by feature concatenation and a lightweight classification module. Experiments were conducted on the NinaPro DB2 dataset using data from 40 subjects and 49 non-rest hand gestures under a repetition-based within-subject evaluation protocol. Repetitions 1, 3, 4, and 6 were used for training, whereas repetitions 2 and 5 were reserved for testing. With causal majority voting over five consecutive predictions, EdgeNetTF achieved a Causal5 accuracy of 87.01% ± 4.30% and a macro-F1 score of 87.41% ± 3.94% under the primary within-subject protocol while requiring only 0.1891 million trainable parameters and approximately 70.01 million FLOPs. Additional cross-subject and external-dataset evaluations revealed substantial inter-subject variability, indicating that subject-independent generalization remains challenging. Overall, the results support lightweight time-frequency representation learning as a computationally efficient approach to multi-class sEMG recognition, while highlighting the need for improved generalization across subjects and populations.