Comparison of Baseline CNN Models for Static Dactyl Recognition in American and Russian Sign Language

Abstract

Advances in digital technologies are enabling solutions to important social problems, such as communication difficulties with people with hearing impairments. The aim of this article is to compare methods for classifying static gesture images. This article presents the results of a comparison of methods for classifying static gesture images using convolutional neural networks based on the LeNet and VGG architectures. A study using datasets of dactylemes from two languages of the French family (ASL and RSL) revealed comparable recognition quality. It is shown that increasing the neural network dimension and the complexity of the feature graph allows for an increase in recognition quality metrics to 0.97-0.99. This suggests a shift in the focus of research to the study and interpretation of feature space, searching for cheremes and related structures.