Deep Learning for Optical Character Recognition of Arabic Text

Optical Character Recognition (OCR) of non-Latin scripts, such as Arabic script
have been investigated over several decades ago. However, the recent advancement
of deep learning has attracted many to improve existing solutions to OCR. Arabic
writing system has its own distinctive style. Words are written from the right to the
left. Features like ligatures, diacritics and vowel markings are commonly included in
writing. Some characters may extend and overlap on top of another characters. This
may cause difficulties when training models using conventional training method.
Therefore, techniques chosen should be able recognize such features. Since Arabic
calligraphy exists in many styles, this study will only focus on recognizing printed
and written khat naskh. Using neural networks technology with the help of enormous
and reliable datasets, models can be trained to get high accuracy and precision in
recognizing the text. In this study, a hybrid neural network model will be built, which
will concern on feature extraction and classification as to encounter difficulties in
OCR of Arabic text.