Transformer-based hand gesture recognition from instantaneous to fused neural decomposition of high-density EMG signals

The box plots show the IQR for each window size that decreases minimally from (\text {CT-HGR})-V1 to (\text {CT-HGR})-V2. 8 represents precision, recall, and F1 score associated with Model (\text {CT-HGR})-V1 for each gesture based on the confusion matrix shown in Fig. More specifically, (\text {CT-HGR})-V3 uses HD-sEMG decomposition to extract microscopic neural drive information from HD-sEMG signals for hand gesture recognition. Then, the final class tokens of (\text {CT-HGR})-V1 and (\text {CT-HGR})-V3 are joined together and fed to a FC layer for final classification. Table 10 compares accuracy and STD for (\text {CT-HGR})-V1, (\text {CT-HGR})-V3 and their fused model for each fold.