Light-Weight CNN-Attention Based Architecture Trained with a Hybrid Objective Function for EMG-Based Human Machine Interfaces

The presented research focuses on Hand Gesture Recognition (HGR) utilizing Surface-Electromyogram (sEMG) signals. This is due to its unique potential for decoding wearable data to interpret human intent for immersion in Mixed Reality (MR) environments. The existing solutions so far rely on complicated and heavy-weighted Deep Neural Networks (DNNs), which have restricted practical application in low-power and resource-constrained wearable systems. In this work, we propose a light-weight hybrid architecture ( $$\text {HDCAM}$$ ) based on Convolutional Neural Network (CNN) and attention mechanism to effectively extract local and global representations of the input. The proposed $$\text {HDCAM}$$ model with 58, 441 parameters reached a new state-of-the-art (SOTA) performance with $$83.54\%$$ and $$82.86\%$$ accuracy on window sizes of 300 ms and 200 ms for classifying 17 hand gestures. The number of parameters to train the proposed $$\text {HDCAM}$$ architecture is $$18.87 \times $$ less than its previous SOTA counterpart. Furthermore, the model is trained based on a hybrid loss function consisting of two-fold: (i) Cross Entropy (CE) loss which focuses on identifying the helpful features to perform the classification objective, and (ii) Supervised Contrastive (SC) loss which assists to learn more robust and generic features by minimizing the ratio of intra-class to inter-class similarity.