Development of an Intelligent System for Recognizing Islamic Religious Visual Signs in the Arabic Language

Duaa Mhnaa, Yaroub Dayoub, Jafar Salman; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2025, pp. 4933-4941

This paper proposes a knowledge-based framework for recognizing the Arabic Islamic religious sign language, developed under data-constrained conditions using a manually constructed and annotated data set. The framework consists of three main components: (1) Integration of motion features derived from hand movements, facial gestures, and body posture using MediaPipe; (2) Transformation of visual signs into text-aligned digital representations; (3) Classification of sign sequences via the SetFit model, a few-shot learning approach, enhanced by semantic retrieval through FAISS to optimize matching efficiency and improve performance. To support the framework, a manually labeled dataset comprising 335 motion representations was created in 11 distinct religious categories. The signs were extracted from long religious video content using precise segmentation to isolate individual sign sequences. The proposed model achieved a classification accuracy of 98.5% on the test set. This study presents a novel resource-efficient approach that combines lightweight classification techniques with semantically aware textual embeddings to address the challenges of recognizing complex religious gesture sequences under limited data conditions. The results open new avenues for artificial intelligence applications in religious education, accessibility, and cultural visual archiving.