Please use this identifier to cite or link to this item:
https://ptsldigital.ukm.my/jspui/handle/123456789/476659
Title: | Quranic optical text recognition using deep learning model |
Authors: | Idris Saleh Ahmed Al-Shiekh (P86689) |
Supervisor: | Masnizah Mohd, Prof. Dr. |
Keywords: | Universiti Kebangsaan Malaysia -- Dissertations Dissertations, Academic -- Malaysia Quran Text recognition Deep learning model |
Issue Date: | 7-Oct-2019 |
Description: | The Holy Quran which is the words of Allah (SWT) verbally revealed to his Prophet Muhammad (SAW) through the angel Jibril, this book is the most significant book for all Muslims around the world. The Calligraphy of the Holy Quran is unique compared to the traditional Arabic written style. One of these characteristics is the presence of diacritics, and it was written by the hand of Arabic calligraphy artist using the Uthmanic script with different spelling rules for some words. In this work, we used the standard Mushaf al Madinah benchmark where there are some rules in writing style, for example, the page should start with the beginning of Ayah and end with the end of Ayah. Following these rules make the words vary in size and paragraphs on different pages. These characteristics making the recognition of the Quranic text more challenging than the normal Arabic text. Most of the resent work on (AOCR) Arabic Optical Character Recognition used the deep learning approach to overcome some of the known issues of the Arabic text like the segmentation, eextraction of hand-crafted feature and tack advantage of the sequence context in the text. the use of the deep learning lead to improve the accuracy of AOCR, but some area still needs more afford to get acceptable accuracy such as the Quranic text where the state-of-the-art systems fails to recognize the Quranic text. This work presents a Quranic Optical Character Recognition (OCR) system based on Convolutional Neural Network (CNN) followed by Recurrent Neural Network (RNN). In this work, six deep learning models were built to study the effect of different representations of the input and output to the accuracy and the performance of the models, also to compare between Long Short-Term Memory (LSTM) and the Gated Recurrent Unit (GRU). A Quranic OCR dataset was built to train these models. The Quranic dataset contains 604 images on page level and 8927 images in text-line level. This dataset is public and free to use for the research community. Moreover, the dataset provides another set of text line image where the diacritics and dots are removed from that set. This work achieves 95.37%,95.28%,91.97%,93.2%,89.05%,89.08 Word Recognition Rate (WRR) and 99%,98% for Character Recognition Rate (CRR) on the six-experiment using the test dataset. The contribution of this work is a Quranic OCR models capable of recognizing the diacritic text of the Quranic image. This work represents a comparison between the LSTM and GRU in the Arabic text recognition domain. A public database has been built to help solve the problem of the lack of public databases for research purposes in the field of Arabic texts recognition, which contain the diacritics and the Uthmanic script and large enough to be used with the deep learning models.,Master of Computer Science |
Pages: | 91 |
Publisher: | UKM, Bangi |
Appears in Collections: | Faculty of Information Science and Technology / Fakulti Teknologi dan Sains Maklumat |
Files in This Item:
File | Description | Size | Format | |
---|---|---|---|---|
ukmvital_123749+SOURCE1+SOURCE1.0.PDF Restricted Access | 2.78 MB | Adobe PDF | View/Open |
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.