A Lightweight Real-Time Recognition System for Static Quranic Fingerspelling in the Malaysian Sign Language Context

Abstract

Accessible Quranic learning for Deaf learners in Malaysia requires recognition of Quran Rasm handshapes used within Malaysian Sign Language (MySL) practice. However, existing approaches provide limited support for real-time feedback on the handshapes performed by learners, and no dedicated public dataset is currently available for Quran Rasm in MySL. This study develops a lightweight real-time recognition application for 29 static Quran Rasm handshapes using MediaPipe hand landmarks and a compact feedforward neural network. Three geometric preprocessing configurations based on wrist normalization, hand mirroring, and scale normalization were evaluated using stratified five-fold cross-validation across multiple model initializations while maintaining the same classifier architecture. The recognition model was integrated into a multithreaded desktop application incorporating temporal stability filtering, interactive quiz generation, correctness feedback, scoring, optional answer hints, and result saving. The best preprocessing configuration achieved a macro-averaged F1-score of 0.9608, while the classifier contained 9,617 parameters with a model size of approximately 42 KB. The completed application was also used by 17 Deaf learners in the intended school environment, demonstrating practical deployment and direct learner interaction. These findings show that lightweight landmark-based recognition can be integrated into a functional Quran Rasm learning application for the Malaysian context.