الفهرس 
IntroductionOnly 14 pages are availabe for public view
 |  | from | 128 |  |  |
| | | from | 128 | | |
Abstract
As the field of handwriting recognition develops, the ultimate goal is to make machines capable of reading any text as accurately as humans, while doing so faster. It is still hard to recognize offline Arabic handwritten words.
Chapter 1 introduces the motivation and challenges related to automatically recognize the offline Arabic handwritten words, which could be divided into two categories: one re- lated to the nature and characteristics of the language, such as cursive nature, ligatures, and overlapping characters; and the other related to the quality of available datasets, such as imbalanced datasets. The goal of this work is to classify each handwritten word so that the handwritten text can be converted to a digital form.
Chapter 2 shows the literature review for solving this problem. First, it presents the general pipeline of the previous classical machine learning algorithms, which required segmenting and manually extracting the word features before the stage of recognition. Also, it clarifies the previous work done by using deep learning techniques such as CNN and its components, RNN and its variants.
In order to accomplish this task Chapter 3 proposed two main approaches: Ensem- bles and transfer learning techniques. The first proposed model is based on averaging the output predictions of three convolutional neural networks, in which each convolu- tional neural network was trained differently using different optimizers. By training three CNNs with different optimization techniques, the output predictions become more diverse, resulting in an improvement in the final output prediction when averaging their predictions. To overcome the problem of insufficient training samples, the second pro- posed model employs the transfer learning technique. This study investigates how deep learning techniques can improve this recognition accuracy to overcome the above chal- lenges.
The results of our study in chapter 4 showed that several deep learning techniques are capable of extracting and learning discriminative features of Arabic words if there are sufficient and equal training samples (balanced datasets). Also, it presents the details of all our experiments and their results related to evaluating different DCNN architectures and our two proposed models.
Chapter 5 gives an interpretive reason for the misclassification of words resulting from our proposed models. Also, it provided a comparative analysis of other proposed models in the literature.
Finally, chapter 6 shows the main conclusions points and our proposal to the future works.