Contribution à la reconnaissance hors ligne de l'écriture arabe manuscrite

Abstract

Offline Arabic Handwritten Text Recognition (HTR) is an important research area. Contributing to its development may have significant and exciting implications for many Arabic-speaking countries, where Arabic is the official spoken and written language. This thesis has examined the Arabic handwritten language, its characteristics, and the challenges of recognizing its written script. We conducted thorough literature reviews on classical and modern deep learning methods for recognizing Arabic script. We delve deep into the resources used to train the Arabic HTR systems, trying to extract all the known databases of Arabic handwritten text. We also investigated the data augmentation approaches used to enrich the handwritten Arabic text databases with new data. Later, we proposed a new deep learning architecture based on Convolutional Recurrent Neural Network (CRNN) to recognize offline Arabic scripts. As a final contribution, we proposed a new data augmentation technique based on Moving Least Squares (MLS), specifically designed to generate new images of Arabic handwritten text to serve as training data for deep learning systems.