Abstract: This study presents the implementation of a convolutional neural network focused on the recognition of hand gestures for this case 3 specific types of gestures using the EMG signals as input which were acquired through the Myo armband device and processed by means of a characteristic map extraction technique which is the power spectral density. The development of this work is divided into 2 phases where the first consists of the acquisition and processing of the electromyographic signals of different users with different arm thickness from which 2 databases were built and the second phase describes the architecture of the convolutional neural network to be used and the training that was performed with each database independently, obtaining two trained networks. Finally, two types of tests are performed, a validation test in which the accuracy of the two trained networks is verified where a accuracy rate of 91.7 and 92.5% was achieved and a real-time behavioral test where the two networks responded adequately, meaning that the use of convolutional neural networks for the recognition of hand gestures by means of electromyographic signals can reach high ranges of accuracy, even greater than 90%.
Javier O. Pinzon Arenas, Robinson Jimenez Moreno and Ruben D. Hernandez Beleno, 2018. Hand Gesture Recognition Using Electromyographic Signals Throw a Deep Convolutional Neural Network. Research Journal of Applied Sciences, 13: 482-490.