الفهرس | Only 14 pages are availabe for public view |
Abstract Optical character recognition (OCR) is a process of converting an image of handwritten or printed text into digitized form. Latin character recognition has been extensively investigated using di{uFB00}erent techniques. However, little work has been done for Arabic OCR due to the challenges that face the Arabic domain. Most of the previous work that has been done in recognition of separate Arabic characters uses hand-crafted features as well as trainable classi{uFB01}er. In this work, we study the power of deep neural networks for Arabic character recognition task. We explore the ability of deep neural networks to learn power features that are invariant to some degree of shift, rotation, scale, geometric distortions, and di{uFB00}erent handwritten styles. We tackle the OCR problem using two approaches. In the {uFB01}rst approach, we propose two deep neural networks models: stacked auto-encoder and convolution neural network. We present a comparative study between the two models. As the deep neural networks need a huge number of samples for training, the previous approach su{uFB00}ers in case of small dataset which leads to our second approach. In this approach, we propose a siamese neural network model for one shot classi{uFB01}cation task. The system can generalize to new data from unseen target classes with high recognition rate without the need for retaining the deep neural network model |