Wearable FSS and Machine Learning: Development and Validation of a Convolutional Neural Network for Recognition of Brazilian Sign Language (Libras)

This study addresses the implementation, training, and testing of a 2 D neural network with three convolutional layers and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">${3} \times {3}$</tex> filters, trained to identify Brazilian Sign Language (Libras) hand configurations. The data used to train the neural network consist of images of the transmission coefficient (S21) graph obtained from an FSS data glove. The training and testing datasets were initially subjected to various image processing techniques to augment the database, resulting in a dataset composed of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">${3, 0 2 4}$</tex> distinct images generated from the original 28 samples. For training and validation, the network achieved accuracies of 95% and 97%, respectively, using the specified architecture. This architecture was determined based on training and testing the model with different configurations. The results demonstrate that the developed machine learning model is suitable for integration into a computational system coupled with a wearable FSS device to identify and classify Libras hand configurations.