A study of feature extraction for Arabic calligraphy characters recognition

Abstract

Optical character recognition (OCR) is one of the widely used pattern recognition systems. However, the research on ancient Arabic writing recognition has suffered from a lack of interest for decades, despite the availability of thousands of historical documents. One of the reasons for this lack of interest is the absence of a standard dataset, which is fundamental for building and evaluating an OCR system. In 2022, we published a database of ancient Arabic words as the only public dataset of characters written in Al-Mojawhar Moroccan calligraphy. Therefore, such a database needs to be studied and evaluated. In this paper, we explored the proposed database and investigated the recognition of Al-Mojawhar Arabic characters. We studied feature extraction by using the most popular descriptors used in Arabic OCR. The studied descriptors were associated with different machine learning classifiers to build recognition models and verify their performance. In order to compare the learned and handcrafted features on the proposed dataset, we proposed a deep convolutional neural network for character recognition. Regarding the complexity of the character shapes, the results obtained were very promising, especially by using the convolutional neural network model, which gave the highest accuracy score.