Improved Gesture Recognition Using Deep Neural Networks on sEMG

Proper classification and detection of hand muscle movements have always been an important aspect for prosthesis control. The muscular movements can be best understood practically by the analysis of non-invasive surface electromyography (sEMG). But the challenge is to differentiate various types of finger movements with high accuracy. For this purpose, we used deep convolutional neural networks to differentiate and classify a total of ten individual and combined finger movements using a two channel sEMG. We used a database of ten participants with two electrodes attached on their forearms performing our preselected movements. Our aim was to increase the maximum classification accuracy using best practices of neural network and evaluating numerous hyperparameters. The final testing accuracy was approximately 98.88 to 100% for our dataset which is significantly higher than many conventional processes.