Title: A comparative study of two deep learning architectures for gesture recognition on ArSL2018 dataset

Authors: Houssem Lahiani; Mondher Frikha

Addresses: Faculty of Sciences of Bizerte, University of Carthage, Tunisia; National School of Electronics and Telecommunications, University of Sfax, Tunisia; Advanced Technologies for Image and Signal Processing (ATISP) Laboratory, Technopark of Sfax, Tunis Road 10 km., P.O. Box 1163 SFAX 3021, Tunisia ' Faculty of Sciences of Bizerte, University of Carthage, Tunisia; National School of Electronics and Telecommunications, University of Sfax, Tunisia; Advanced Technologies for Image and Signal Processing (ATISP) Laboratory, Technopark of Sfax, Tunis Road 10 km., P.O. Box 1163 SFAX 3021, Tunisia

Abstract: This study presents a comparative evaluation of two convolutional neural network (CNN) architectures for Arabic sign language recognition using the ArSL2018 dataset, which contains 54,049 images across 32 alphabet classes. The objective is to identify an efficient model that can support assistive technologies for the Arabic-speaking deaf community. The first approach applies a pre-trained MobileNetV2 as a feature extractor with a fully connected layer, while the second extends MobileNetV2 by adding convolutional and pooling layers. Performance was assessed through accuracy, precision, recall, and F1-score. Results indicate that the pre-trained MobileNetV2 configuration achieved superior accuracy (95%) compared to the extended model (93.85%), with per-class performance ranging from 82.91% to 99.10%. These outcomes highlight the potential of lightweight CNN architectures for real-world deployment, offering both effectiveness and efficiency in enhancing communication accessibility through gesture-based interaction technologies.

Keywords: convolutional neural network; CNN; ArASL; MobileNzetV2; HMI.

DOI: 10.1504/IJAISC.2025.151125

International Journal of Artificial Intelligence and Soft Computing, 2025 Vol.9 No.3, pp.209 - 232

Received: 28 Aug 2024
Accepted: 03 Sep 2025
Published online: 14 Jan 2026 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article