Handwritten Arabic Character Recognition: Comparison of Conventional Machine Learning and Deep Learning Approaches

Abstract

Over the last decades, automatic handwriting recognition has received a lot of attention, as it is a crucial component for many applications in various fields. Research for this issue has focused on handwriting recognition in Latin languages and fewer studies have been dedicated to the Arabic language. In this paper, we propose and compare two approaches to classifying Arabic characters. The first is based on conventional machine learning using the SVM classifier by comparing different sets of features, most commonly used in the pattern recognition field. The second is based on deep learning by testing different CNN (convolutional neural networks) architectures, which brings a self-characterization of Arabic features. In this context, a new fast and simplified CNN architecture is proposed. We also test different transfer learning strategies on two versions of the OIHACDB dataset and the AIA9K dataset proposed in the literature. In the experimental section, we show that the proposed CNN model achieves accuracies of 94.7%, 98.3%, and 95.2% on the test set of the three databases OIHACDB-28, OIHACDB-40, and AIA9K respectively. Our experiments enrich the tests already carried out on these datasets and show good results in comparison with the literature.