Recognition of Libras Static Alphabet with Myo™ and Multi-Layer Perceptron

Abstract

A Sign Language is a structured set of corporal gestures used as a communication system, which uses movements of the arm, hand, forearm, facial expressions, and lips movements to ease the communication among deaf and/or hearing people. In Brazil, the official Sign Language is called Libras. This work presents the recognition of static alphabet of Libras (20 letters) using the armband Myo™ and a Multi-Layer Perceptron. Myo™ captures Electromyography signals from forearm and these signals are used to classification. The data were acquired from one male subject, 42 times for each gesture. The signals were segmented in periods of 750 ms using onset technique and 10 features were extract from these segments. The built MLP has one hidden layer, one input layer, and one output layer, trained by the backpropagation algorithm. The number of neurons in hidden layer was tested from 10 to 300 and the best approximation for MLP was 230 neurons. The classification has an accuracy of 91.3 ± 0.5% in training and 81.6 ± 0.9 in the test. Finally, the gestures presented accuracies above 80%, except the gestures ‘L’, ‘R’, and ‘W’.