Abstract
Sign languages are essential for communication among over 430 million deaf and hard-of-hearing individuals worldwide. However, recognizing Arabic Sign Language (ArSL) in real-world settings remains challenging due to issues like background noise, lighting variations, and hand occlusions. These limitations hinder the effectiveness of existing systems in applications such as assistive technologies and education. To tackle these challenges, we propose ASLDetect, a new model for ArSL recognition that leverages ResNet for feature extraction and a U-Net-based architecture for accurate gesture segmentation. Our method includes preprocessing steps like resizing images to 64 × 64 pixels, normalization, and selective augmentation to improve robustness in diverse environments. We evaluated ASLDetect on two datasets: ArASL2018, which features plain backgrounds, and ArASL2021, which includes more complex and diverse environments. On ArASL2018, ASLDetect achieved an accuracy of 99.35%, surpassing ResNet34 (99.08%), T-SignSys (97.92%), and UrSL-CNN (0.98%). For ArASL2021, we applied transfer learning from our ArASL2018-trained model, significantly improving performance and reaching 86.84% accuracy-outperforming ResNet34 (82.5%), T-SignSys (58.98%), and UrSL-CNN (0.49%). These results highlight ASLDetect's accuracy, robustness, and adaptability.