Intelligent ResNet-18 based Approach for Recognizing and Assessing Arabic Children's Handwriting

Studies tackling handwriting recognition and its applications using deep learning have been promoted by developing advanced machine learning techniques. Yet, a shortage in research that serves the Arabic language and helps develop teaching and learning processes still exists. Moreover, COVID-19 pandemic affected the education system considerably in many countries and yielded an immediate shift to distance learning and extensive use of e-learning tools. An intelligent system was proposed and used in this paper to recognize isolated Arabic handwritten characters. Particularly, pre-trained CNN models were exploited and fine-tuned to meet the requirements of the considered application. Specifically, the designed system automatically supports teaching Arabic letters and evaluating children's writing skills. The Arabic Handwritten Character Dataset (AHCD) was used to train the models built upon ResNet-18 and assess the overall system performance. Furthermore, several models were investigated using various hyper-parameter settings in order to determine the most accurate one. The best model with the highest accuracy rate of 99% was used and integrated into the proposed system to recognize the Arabic alphabets.