Arabic Object Character Recognition (OCR) has been an under-researched topic. It faces many difficulties as the characters change shape depending on the position within a word. The words can have different prefixes and suffixes as a result of conjugation. This forces the language model to be trained with a large amount of data to be able to train to recognize all of these different forms. This is difficult because the amount of Arabic-based texts on the internet is small. This work uses traditional Recurrent Neural Networks in text extraction of specific fonts of the Arabic script. We present results to show the efficacy of using domain specific models, alongside font classifiers, in comparison to deep neural networks that are more generic to multi-font and multi-lingual models. We found that domain-specific models outperform other open-source solutions on datasets that contain specific font styles by a five percentage points in some fonts. Having a system that redirect images to font-specific OCR models, using a CNN-based Classifier, generates superior results to the generic approach.
Advisor
Abstract
Publication Type
Publication Year
Subject
Computer Science
Other Contributors