Increase the effectiveness of the Arabic text-to-image generation task

Résumé

Abstract The Processing of Arabic text to generate images is a challenging task. Because the Arabic language belongs to the Semitic group of languages and it is less discovered than English in the Artificial Intelligent community. It is also a right-to-left language that needs special natural language tools to process. On the other hand, efficient text-to-image generation architectures have been built by the Generative Adversarial Network methodology, which had shown incredible results in so many tasks. In this paper, because the text is the first domain of the text-to-image generation task, we increase the effectiveness of Arabic text-to-image generation by implementing tree approaches. Firstly, We fuse the AraBERT with DF-GAN by injecting the AraBERT sentence vector into the DF-GAN generator and discriminator. Secondly, we fuse a sample text transformer with a DF-GANs generator and discriminator to overcome the out-of-vocabulary limitation. Thirdly, we keep the training process and the text transformer, and we add a learning mask predictor to the architecture to predict a mask, which is used in affine transformation parameters to fuse image and text deeper, and we also train the architecture with the DAMSM loss function to increase the stability in the training phase. The experimental results on CUB and Oxford-flower datasets demonstrate the superiority of the third approach in Arabic text-to-image generation tasks.

Citer ce document

Accès au document

Ce document est hébergé sur son dépôt institutionnel d'origine.

Auteur(s)

Statistiques

Consultations : 1

Téléchargements : 0