ArSL-CNN a convolutional neural network for Arabic sign language gesture recognition

Abstract

Sign language (SL) is a visual language means of communication for people who are Deaf or have hearing impairments. In Arabic-speaking countries, there are many Arabic sign languages (ArSL) and these use the same alphabets. This study proposes ArSL-CNN, a deep learning model that is based on a convolutional neural network (CNN) for translating Arabic SL (ArSL). Experiments were performed using a large ArSL dataset (ArSL2018) that contains 54049 images of 32 sign language gestures, collected from forty participants. The results of the first experiments with the ArSL-CNN model returned a train and test accuracy of 98.80% and 96.59%, respectively. The results also revealed the impact of imbalanced data on model accuracy. For the second set of experiments, various re-sampling methods were applied to the dataset. Results revealed that applying the synthetic minority oversampling technique (SMOTE) improved the overall test accuracy from 96.59% to 97.29%, yielding a statistically signicant improvement in test accuracy (p=0.016,  α<0=05). The proposed ArSL-CNN model can be trained on a variety of Arabic sign languages and reduce the communication barriers encountered by Deaf communities in Arabic-speaking countries.