Abstract

RODRIGUEZ VILLARROEL, Juan Pablo; PONCE DE LEON ESPINOZA, Nicolás  and  ARTEAGA SABJA, Wendoline. Multilayer and convolutional neural networks for Bolivian Sign Language recognition: an empirical evaluation. RevActaNova. [online]. 2021, vol.10, n.1, pp.22-41. ISSN 1683-0789.

The deaf community is a social stratum with lots of struggles in daily life, chiefly cause for communication difficulties with the general public. Although each country has its sign language, which is the case of Bolivian Sign Language(BSL). However, only few people know it. Different approaches have been proposed to perform gesture recognitions and help people to translate sign language to a particular language, including neural networks. However, little is known about the effectiveness of the neural networks to detect Bolivian Sign Language (BSL). This paper proposes and evaluates the use of two neural network techniques, multilayer (MLP) and convolutional(CNN), to recognize Bolivian Sign Language. Our approach takes as input the most significant frames from a video using a motion-based algorithm and applying a border detection algorithm in the selected frames. We present an experiment on which we evaluate these techniques using 60 videos of four basic BSL phrases. As a result, we found that MLP has an accuracy which ranges between 65% and 88%, and CNN ranges from 95% and 99%, depending of number of neurons and internal layers used.

Keywords : multilayer neural network; convolutional neural networks; computer vision; sign language recognition; BSL.

        · abstract in Spanish     · text in English     · English ( pdf )