Uso de técnicas de aprendizagem profunda na classificação de configurações de mão de língua de sinais

Abstract

This work presents a method to classify Brazilian sign language hand configurations using convolutional neural networks. The network architectures used were selected based on a systematic bibliographic research. Several experiments were done using different values of hyperparameters aiming to obtain the best fit the classification task. The models training was carried out for 500 epochs using three different architectures and two regularization techniques (dropout and L2). LibrasImage, a data set of hand configurations depth images was used in the training and testing steps of the models. The models were analyzed with respect to the accuracy, sensitivity, area under the ROC curve and error rate for each hand configuration. The best result obtained was an accuracy of 97.98%. This result shows that the use of convolutional neural network improves the classification of Brazilian sign language hand configurations in relation to the method that uses the k-nearest neighbor classifier, that was tested with the same dataset. The difference in performance between the two methods was statistically significant by Pearson chi-square test.