Abstract Arabic Sign Language (ArSL) is an important medium of communication for hearing and deaf communities. Recognition of ArSL is difficult, especially in signer-independent systems. Mostly studies deal with ArSL recognition from images or videos and dependent on the signers. To simulate the real world, we must introduce ArSL recognition models based on videos with independent signers to recognize signature style diversity. This paper introduced an ArSL-TGRU model to recognize Arabic sign language from videos using a hybrid transformer and Gated Recurrent Unit (GRU). Four sequential transformer layers improve temporal and spatial abstraction, while two simultaneous GRU modules capture sequential dynamics. To capture complex dynamic movements, this network reduces overfitting and models long-range relationships via attention techniques and regularization. The proposed model applied on a custom dataset extracted from KArSL’s. The custom dataset contains 60,918 videos for 136 signs, divided into 6 categories of signs (letters, numbers, words-family, words-feeling, words-job, and sentences) by three independent signers. From the experiments, the proposed model outperformed six common architectures in several categories with different signers. The proposed model’s ability to understand spatial and temporal connections, adapt to signer variances, and achieve robust generalization across various linguistic forms.
Nasef et al. (Sat,) studied this question.