Optical Recognition System of Non-Dotted Arabic Letters

Abstract

Writing with Arabic characters is a very sensitive operation because some characters are vertically nested and written in multiple ways. The infinite variety of font styles and sizes poses many challenges when designing Arabic recognition systems. In this article, we used deep learning to build an Arabic recognition system that could recognize dotted letters. Our dataset contains 3000 color images for ten letters ( م ,و ,ل ,ع , ,ص ,س,ح ,و ,ط ,ر د ). The training stage was achieved through 2250 images (10 classes) of this data, and 750 images were used for the validation stage. The classification accuracy of our proposed model is 98.67 percent. Experimental results show that Arabic font styles (Diwani, Hacent Dalal, Times New Roman, B Tanabe, and Kofi) have the best classification accuracy, so they are the preferred styles to use in Arabic writing texts. The font sizes 28, 36, 48, and 72 are the letters that have been classified with 100% rate accuracy using our trained model, so they are the font styles preferred to use in writing Arabic texts.