Développement d’un système de reconnaissance de caractères arabes manuscrits

Abstract

The field of automatic handwriting recognition (printed or handwritten) is important because of its versatility. Therefore, the objective of our graduation project is to set up a system for identifying Arabic script from digitized images (offline system). Where, we rely on Convolutional Recurrent Neural Network (CRNN), which combines Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN). Whereas, CNN supports deep computer vision and RNN processes natural language. Additionally, we use connectionist temporal classification (CTC) for encoding. Moreover, we used the most known databases (IFN/ENIT), (ADAB) and we added 714 images of characters to improve our results. After applying the CRNN, we trained and validated our model, and we obtained very interesting results around 96% on the recognition of Arabic words with an error rate of 0.77%.