Reconhecimento de Libras em frames estáticos de vídeos utilizando CNN e técnicas de pré-processamento de imagens.

Abstract

Although Brazilian sign language (Libras) was recognized as an official language of Brazil in 2002 [1], legal measures that regulate and require the offer of Libras teaching in schools to some degree were only reversed in a bill in 2019 [2]. Resulting in a lack of contact of people without hearing problems with Libras, and combined with the finding of the World Federation of the Deaf (WFD) that about 80% of the deaf in the world have problems understanding the written languages of their respective countries [3], generates social isolation with people who depends on the use of hand signals to communicate. In this context, there is the field of recognition of signal languages (RLS), which proposes to create technological interfaces that can act on the described problem. This work uses static video frames extracted from the MINDS-Libras dataset [17] to analyze the impact of using image preprocessing methods in the training of a Convolutional Neural Network (CNN), to obtain a model capable of classifying 20 different signals of Pounds efficiently. In the end, the proposed method reached an average accuracy of 91.08% in the data set used.