Advanced Sign Language Recognition Using Deep Learning: A Study on Arabic Sign Language (ArSL) with VGGNet and ResNet50 Models

Sign language is a vital communication tool for individuals with hearing impairments and those who do not speak the language of the host country. Despite its importance, sign language is not widely known and lacks universal patterns, with each country having its own unique set of signs influenced by local customs and traditions. This complexity presents significant challenges for computer vision-based recognition systems. Currently, extensive research is being conducted to address the issues of sign language translation using advanced technologies. This paper introduces two novel datasets comprising 54,049 images of the ArSL-2018 (Arabic Sign Language) alphabet captured by over 40 contributors, alongside 15,200 images from proprietary datasets and digits sign language datasets. The images are meticulously categorized into 32 classes representing standard Arabic characters and processed through classification, normalization, and detection procedures using VGGNet model and ResNet50, designed to enhance the accurate recognition of hand gestures in real-time, even under complex background conditions. By leveraging deep learning and artificial intelligence, this system aims to bridge the communication gap between deaf and hearing individuals. The investigation delves into the comparison of the outcomes derived from two distinct training and testing regimes, one without fine-tuning and the other incorporating fine-tuning enhancements. The remarkable levels of precision attained in contrast to prior research affirm the efficacy of this methodology. Particularly, the dataset encompassing ArSL alphabets exhibited accuracy rates of 99.05\%, 99.99\%, and 98.50\% when employing VGG16, VGG19, and ResNet50 Models, respectively, thus underscoring the efficiency of the proposed approach for hand gesture recognition tasks.