A Few-Shot Learning-Based Domain-Adaptive Lightweight CNN for Robust Hand Gesture Recognition in Real-Time Applications

Surface electromyogram-based hand gesture recognition (sEMG) has promising prospects in human-machine interaction applications, including the neural rehabilitation and assistive technologies. However, this technique also has several disadvantages, such as poor generalization across subj ects and days, demanding a large amount of data for calibration, and computational complexity that does not enable the real-time deployment of the system. The proposed work incorporates a novel few-shot learning-based domain-adaptive lightweight CNN that is specifically designed to be used in robust and efficient hand gesture recognition. This minimizes the dependency on the labeled data of other new subj ects or days, hence allowing better adaptability in real-world scenarios. The proposed lightweight CNN architecture adapts well to wearable devices with real-time performance but high accuracy. Another would be adding domain-adaptive feature learning to account for variations between inter-subject and inter-day domains while preserving the accuracy of classification and improving model generalization. The proposed framework is able to achieve competitive accuracy using a small calibration set in gesture-based human-machine interfaces continuously and broadly. Results on high-density sEM G datasets have shown that our model achieves a good balance between the accuracy (95.5%) and efficiency on experimental results and holds promise for providing a reliable solution for real-time cross-domain hand gesture recognition applications.