الفهرس 
IntroductionOnly 14 pages are availabe for public view
 |  | from | 128 |  |  |
| | | from | 128 | | |
Abstract
Offline Arabic handwritten word recognition is still a challenging task. Many deep learn- ing approaches perform admirably on this task if the lexicon size is not too large and the number of training samples is sufficient for the training process. The challenges of recognizing offline Arabic handwritten words could be divided into two categories: one related 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 aim of our work is to investigate the ability of different deep learning techniques to improve recognition accuracy and overcome the above challenges. In order to accomplish this task, we used two main approaches: Ensembles and transfer learning techniques. The results of our study showed that several deep learning models are capable of ex- tracting and learning discriminative features of Arabic words if there are sufficient and equal training samples (balanced datasets).
Compared to other state-of-the-art methods, we proposed two models which achieve recognition accuracy of up to 96.11 %. The first proposed model is an ensemble model based on averaging the output predictions of three convolutional neural networks, in which each convolutional neural network was trained differently using different optimiz- ers. The overall recognition accuracy of our proposed model is 93.37%, while the best CNN model achieved an accuracy of 92.7%5.
The second proposed approach is a two stages transfer learning model that is able to overcome the problem of insufficient training samples. We demonstrate that using ImageNet as a source dataset improves the recognition accuracy of the ten frequently misclassified words in the IFN/ENIT dataset by 14%, while our proposed approach re- sults in an increase of 35.45%. Our overall recognition accuracy was 96.11%, which is nearly a 2.5% improvement over other state-of-the-art methods.
In the experiments, two offline datasets were used: the AlexU-W dataset, which con- tains a sufficient number of simple words, and the IFN/ENIT dataset, which is more challenging.