A DEEP LEARNING APPROACH USING RESNET-50 FOR ARABIC SIGN LANGUAGE RECOGNITION

Résumé

Abstract. Arabic Sign Language (ArSL) serves as a vital communication medium for the deaf and hard-of-hearing community within Arabic-speaking
regions. However, significant communication barriers often exist between signers and non-signers, limiting access and inclusion. This paper investigates the effectiveness of the ResNet-50 architecture for ArSL recognition. We employ a transfer learning methodology, fine-tuning a ResNet-50 model pre-trained on ImageNet using the RGB Arabic Alphabets Sign Language Dataset containing
16,000 images across 32 classes representing the Arabic alphabet. Standard image preprocessing and data augmentation techniques are utilized to enhance model robustness. The study achieves 96% recognition accuracy for static ArSL signs, evaluated using standard classification metrics. This work provides a rigorous evaluation of ResNet-50 within the ArSL context, reinforcing the potential of transfer learning for developing practical assistive communication technologies and promoting greater inclusivity for the Arabic deaf community