Comparação de algoritmos de reconhecimento de gestos aplicados à sinais estáticos de Libras

Abstract

Brazilian Sign Language (BSL) has been created in order to cope with a necessity of a non-verbal communication for the deafs, which during a long time were indoctrinated to learn the Brazilian Portuguese as their first language. Nowadays, the BSL is the Brazil’s second official language and first deaf’s language, as well as the Portuguese for the listener. Nevertheless, even with large recognition, the Brazil’s second official language is not known by the majority of the Brazilian population. The inclusion process aims to allow equality for the impaired, such that the deficiency does not become an impediment factor for living together in society. With the technology arrival and the Artificial Inteligence (AI) advances, it was created technologic artifices to allow inclusion. In the AI, the pattern recognition is one of more approached subthemes in the present, and it is widely applied for the gesture classification of many sign languages in literature. This research has, as key task, the identification of the hands that form a certain BSL gesture and, thus, the recognition of the class it belongs to. Based on American Sign Language (ASL) classification, the Feature Fusion-based Convolutional Neural Network (FFCNN), an extended network from Convolutional Neural Network (CNN), obtained the best accuracy in comparison to other networks, such as Visual Geometry Group (VGG). Therefore, based on this scenario, this work applies the FFCNN to BSL static gestures to verify whether the FFCNN obtain the best accuracy as well as obtained in ASL or not. In order to achieve the goal, this work compares three classifiers: the Visual Geometry Group (VGG), a CNN with variation of 13 and 16 layers, the FFCNN, and a Multi Layer Perceptron network used in recognition of BSL static gestures in literature. The algorithms were applied in a BSL dataset with 9,600 images of 40 signals. The results demonstrate that VGG with 16 layers obtained the best accuracy regarding the described models in this work, corresponding to 99,45%.