STMS-Net: Spatial Temporal Multi-Spectral Network for sEMG Gesture Recognition

Accurate gestures recognition of sEMG plays an important role in the field of human-robot interaction. Currently, researchers commonly extracted feature from sEMG signals though spatial or temporal views. This paper introduces a novel Spatial Temporal Multi-Spectral Network (STMS-Net) to improve the recognition capability by leveraging the spatial, temporal and spectral features of sEMG signals. First of all, sEMG signal is filtered by a filter bank to obtain a multi-spectral data, then it is segmented into several fragments. Afterwards, for each fragment, a Multilayer Perceptron (MLP) Combining Max Pooling Network (MCMP-Net) is employed to extract the local spatial features. After that, the spatial features undergo two self-attention based modules to aware both the temporal and spectral interdependence within them, in those two modules, the spatial features are regarded as tokens of attention layer from spatial or spectral views, and dimension rearrangement is employed to streamline the network operations. At last, MLP and Softmax layer are operated as classification head. We evaluate our model on eight benchmark datasets: NinaPro DB1, DB2, DB4, DB5, CapgMyo DB-a, DB-b, and DB-c, and BandMyo, the accuracy on them are 91.9%, 90.0%, 90.1%, 96.1%, 94.1%, 93.8%, 95.1%, 76.9%, respectively, the results show that STMS-Net achieves or exceeds state-of-art performance.