Generative Adversarial Networks (GAN) for Arabic Calligraphy

Abstract

Arabic calligraphy is one of the most aesthetic art forms in the world due to its variety and long history. However, generating calligraphic style is mainly done by human expert calligrapher (also known as Khattat) and has not been carried out by machine learning techniques. Generative adversarial networks (GAN) are deep learning tools that achieved outstanding results in the field of style transfer and generation. In this paper, various GAN architectures were investigated such as CycleGAN, Pix2pix, and deep convolutional generative adversarial networks (DCGAN) within Arabic calligraphy in two aspects: generation and style transfer. The results show that CycleGAN can transfer skeleton letters to both Naskh and Thulth styles, Pix2Pix can denoise the calligraphy papers, and DCGAN can generate realistic Arabic calligraphy letters. The proposed approaches are applicable for other calligraphy styles besides Naskh and Thulth. Finally, the models are evaluated qualitatively using a preference judgment technique survey.