Résumé: The Arabic language is characterized by a great diversity in its writing styles in terms of the shape, size, and percentage of inclination and methods of drawing. Despite this great diversity, the Arabic styles are also characterized by a huge similarity that makes it difficult for traditional methods of machine learning to overcome with Arabic manuscripts. In this paper we present a Morphological Gradient Convolutional Neural Network to Classify the Arabic styles (MG-CNN). The model is a combination of two methods: a morphological gradient to detect images’ contours and a convolutional neural network to extract images features and classify them. Due to the absence of Arabic styles dataset, we created an image database from the book “Teach Yourself Arabic styles: Naskh, Rokaa, Farissi, Tolot, Diwani” (Mehdi, Teach yourself arabic styles: Naskh, Rokaa, Farissi, Tolot, Diwani, 2005) and then we use augmentation methods to increase number of images while preserving the characteristics of each style. Our architecture gives a high accuracy of 100% for the training dataset and 99.50% for the validation dataset.

Langue: Anglais