Automated Arabic Sign Language Recognition System Based on Deep Transfer Learning

World Health Organization has reported that about 5% of the world's population is hearing-impaired. Automated sign language recognition system interfaces are classified into direct-device, vision-based and hybrid-based interfaces. Deep learning methodologies have been proven as an excellent tool for several automated computer vision systems. Moreover, deep learning overcame several difficulties existed inside traditional computer vision systems, A crucial need is found to provide deaf people with easy deep learning methods to interact with other people. In this paper, we propose a robust recognition system for Arabic sign language based on deep transfer learning. We employ transfer learning based on fine-tuning of existed pre-trained networks. In addition, we employ the data-augmentation to avoid overfitting and increase overall system performance. Several networks architectures have been examined for our target recognition task. We have also investigated the performance of residual networks versus plain networks. During our experiments, we employed Arabic sign language (ArSL2018) public dataset that consists of 54,049 images with 32 class. The overall system accuracy achieved by the proposed one is 99.52% and 99.5% sensitivity based on ResNet18 Architecture with data augmentation benefits. A powerful Arabic sign language recognition system based on deep learning theory is proposed which can be employed later in several automated sign language recognition tasks.