Developing arabic sign language recognition system

This work proposes a complete solution for Arabic Sign Language (ArSL) identification using both conventional machine learning, and deep learning techniques. on a dataset regarding 7,857 ArSL pictures covering 31 classes, we assessed several feature extraction methods (HOG, SIFT, SURF, LBP) using conventional classifiers (SVM, Random Forest) versus deep learning models CNN, MobileNet+Transformer. with SVM-RBF, our investigations found, that HOG features greatly outperformed other extraction techniques, SIFT, SURF, and correspondingly. this shows how well appropriate feature engineering may significantly enhance conventional model performance. Although Google Colab has hardware restrictions, our MobileNet+Transformer model attained 76% accuracy subsequent to only 20 epochs, and astonishingly reached 98% subsequent to longer training to 100 epochs. for various sign languages as American (99.22%), Irish (>99%), and German (97.5%), its performance beats state-of- the-modern systems. We used active memory management, resolution scaling, and batch size reduction among other technological tweaks to go beyond memory limits. Our results demonstrate, that deep learning models automatically learn hierarchical representations, hence removing human engineering even within the event, that conventional approaches need complex feature extraction. This work provides insightful analysis regarding ArSL recognition, and shows that, with suitable approaches, performance on par with international standards happens to be possible even with constrained processing capacity.