GEAC: Generating and Evaluating Handwritten Arabic Characters Using Generative Adversarial Networks
Tahani Alkhodidi, Lamia Aljoudi, Aisha Fallatah, Aya Bashy, Nahed Ali, Norah Alqahtani, Nujood Almajnooni, Ahad Allhabi, Telawa Albarakati, Tarik Alafif, Sreela Sasi · 2021
Generative Adversarial Network (GAN) has made a breakthrough and great success in many research areas in computer vision. Different GANs generate different outputs. In this research work, we apply different GANs to generate handwritten Arabic characters. A basic GAN, Vanilla GAN, Deep Convolutional GAN (DCGAN), Bidirectional GAN (BiGAN), and Wasserstein GAN (WGAN) are used. Then, the results of the generated images are evaluated using native-Arabic human and Fréchet Inception Distance (FID). The qualitative and quantitative results are provided for the images generation and evaluation. In experimental evaluation, WGAN achieves better results in FID with a value of 96.007. On the other hand, DCGAN achieves better results in native-Arabic human evaluation with a value of 35%.