Deep Arabic document layout analysis

Abstract

Document layout analysis (DLA) is an essential step for Optical Character Recognition Systems (OCR). The text of the document fed to the OCR must be extracted first and isolated from images if exist. The DLA task is difficult as there is no fixed layout for all documents, but instead, there are several layouts. There are various approaches for DLA for various different languages. In this paper, some of the previous techniques used in this field will be listed and then we will discuss the proposed method that depends on deep learning for documents' text localization. We used Arabic Printed Text Image database (APTI [19]), ImageNet [18] and a dataset collected from different Arabic newspapers for training and evaluation.