Document Details
|
Document Type |
: |
Thesis |
|
Document Title |
: |
ARABIC MANUSCRIPT IMAGES RETRIEVAL USING DEEP LEARNING استرجاع صور المخطوطات العربية باستخدام التعلم العميق |
|
Subject |
: |
Faculty of Computing and Information Technology |
|
Document Language |
: |
Arabic |
|
Abstract |
: |
The ancient Arabic manuscripts are valuable pieces of historical information that reflect the education, culture, society, and tradition during specific time periods. Therefore, due to their significance and substantial role in enriching the valuable historical information, this study aims to collect the ancient Arabic manuscripts in a dataset. Then, classify its images to be able to retrieve the ranked top similar images to a user query image accurately and instantaneously. The retrieval system can be according to different search criteria. We satisfied the retrieval according to the manuscript, author, and the calligraphy. The automatic extraction and classification according to the most distinguishable features, is a crucial step to detect the similarities among images successfully. Considering that, the historical Arabic manuscripts are text-based images. Thus, it is important to extract the text from the images and to retrieve them according to their textual features. This step is performed through developing an optimized bidirectional LSTM deep learning model including attention and batch normalization layers. Then,the similarities among the textual contents measured using three different distance metrics named: Manhattan, Euclidean, and Cosine. The manuscripts’ images are also not purely textual. Instead, they include handwritten signatures, drawings, figures, tables, side-notes, …etc. Thereby, it is necessary to consider the non-textual parts within the images and to retrieve them according to their visual features. To accomplish this, we transferred learning from four pre-trained convolutional neural networks named: MobileNetV1, DenseNet201, ResNet50, and VGG19. The Siamese deep learning model along with the three distance metrics tested for measuring the similarities among the images and retrieve them. The most accurate visual and textual deep learning models were fused at three different fusion-levels named: decision-level, features-level, and score-le |
|
Supervisor |
: |
Dr. Lamiaa Elrefaei |
|
Thesis Type |
: |
Doctorate Thesis |
|
Publishing Year |
: |
1442 AH
2020 AD |
|
Added Date |
: |
Thursday, August 27, 2020 |
|
Researchers
| منال محمود خياط | Khayyat, Manal Mahmoud | Researcher | Doctorate | |
|
Back To Researches Page
|