Arabic Sign Language Alphabet Classification via Transfer Learning

The integration of artificial intelligence (AI) has addressed the challenges associated with communication with the deaf community, which requires proficiency in various sign languages. This research paper presents the RGB Arabic Alphabet Sign Language (ArASL) dataset, the first publicly available high-quality RGB dataset. The dataset consists of 7,856 meticulously labeled RGB images representing the Arabic sign language alphabets. Its primary objective is to facilitate the development of practical Arabic sign language classification models. The dataset was carefully compiled with the participation of over 200 individuals, considering factors such as lighting conditions, backgrounds, image orientations, sizes, and resolutions. Domain experts ensured the dataset’s reliability through rigorous validation and filtering. Four models were trained using the ArASL dataset, with RESNET18 achieving the highest accuracy of 96.77%. The accessibility of ArASL on Kaggle encourages its use by researchers and practitioners, making it a valuable resource in the field ( https://www.kaggle.com/datasets/muhammadalbrham/rgb-arabic-alphabets-sign-language-dataset ).