Deep CNN and SimpleRNN Architectures for Bi-Handed Arabic Alphabet Hand-Gesture Recognition

This research introduces ArASL-EduNet for bi- handed static Arabic alphabet hand-gesture recognition from RGB images. The framework is built on a curated dataset of 11,200 labeled images covering 56 balanced gesture classes derived from the 28 Arabic letters, with distinct left- and right- hand realizations for each letter. Four deep learning models were evaluated under the same protocol: CNN, MobileNetV2, EfficientNetB0, and SimpleRNN. Results show that CNN achieves the best performance with 99.6% validation accuracy, 99.28% test accuracy, a macro-AUC of 0.995, and 0.6 ms average inference latency using about 1.1M parameters. EfficientNetB0 follows with 99.1% accuracy, and MobileNetV2 reaches 98.7%. In contrast, SimpleRNN achieves 73.4% validation accuracy, 72.06% test accuracy, and 1.9 ms latency with about 11K parameters. These results confirm that CNN-based spatial feature learning is more effective than reshaped sequence modeling for static Arabic alphabet gesture recognition. The main contribution is the ArASLEduNet dataset and an empirical benchmark for bi-handed Arabic alphabet recognition. Keywords—Arabic Sign Language, Hand Gesture Recognition, Convolutional Neural Networks, Recurrent Neural Networks, Image Classification.