Comparative Analysis of Machine Learning and Deep Learning Techniques for Hand Gesture Recognition Using Surface Electromyography

Hand gesture recognition for the control of rehabilitation and assistive technologies is still a field to be explored. This study investigated the effectiveness of different methods for classifying hand gestures using surface electromyography (sEMG) signals collected from the upper Thalmic Myo armband in experiment two of the Ninapro 5 database. Initially, traditional Machine Learning (ML) techniques were applied, an extraction of temporary characteristics was made, and dimensionality reduction through Principal Component Analysis (PCA) and the Kruskal-Wallis test, followed by four classical classifiers, Decision Trees (DT), K-Nearest Neighbor (KNN), Bayes-Gaussian (NBG) and Support Vector Machine (SVM). However, the results were unsatisfactory, with a maximum accuracy of only 0.10. Subsequently, Deep Learning (DL) techniques were adopted, where the sEMG signals were transformed into images through the Fourier Transform, and although they showed a significant improvement, with a training accuracy of approximately 0.90, the accuracy in validation and testing remained around 0.50. This indicates possible overtraining issues and a lack of differentiation between classes of images. Despite these challenges, DL proved superior to ML in classifying hand gestures from sEMG signals. It is concluded that further research and improvement of signal-to-image transformation techniques are needed to achieve more accurate and feasible classification in real-time applications.