Hybrid Meta-Learning for End-to-End Arabic Sign Language Recognition and Translation using ResNet50"

BackgroundEffective communication between the Deaf community and the hearing society relies on accurate recognition of Arabic Sign Language (ArSL). By fusing speech synthesis with visual recognition, deep learning (DL) has lately powered major developments in SL translationMethods This study proposes an end-to-end deep learning framework that translates isolated ArSL signs into written and spoken language. The framework integrates a vision-based recognition module with text-to-speech synthesis, enabling signs to be recognized at the character level, reconstructed into words, and converted into speech. Feature extraction was conducted using pre-trained CNN architectures, VGG16 and ResNet50, with comparative analysis showing that ResNet50 achieved superior performance. To enhance generalization under few-shot learning conditions, Model-Agnostic Meta-Learning (MAML) and Prototypical Networks (PNW) were applied to the extracted features. ResultsExperimental results demonstrate that the ResNet50-based approach achieved 97.84% accuracy across 31 ArSL classes represented in a 2048-dimensional feature space, while also showing robustness in recognizing previously unseen signs. ConclusionThese findings highlight the effectiveness of combining ResNet50 with meta-learning, providing a solid foundation for future multimodal, speech-enabled assistive technologies to support the Deaf and hard-of-hearing community.