Details

Autor(en) / Beteiligte

• Wang, Zekang
• Liu, Li
• Zhang, Huaxiang
• Ma, Yue
• Cui, Huailei
• Chen, Yuan
• Kong, Haoran

Titel

Generative Adversarial Networks Based on Dynamic Word-Level Update for Text-to-Image Synthesis

Ist Teil von

• 2022 7th International Conference on Image, Vision and Computing (ICIVC), 2022, p.641-647

Ort / Verlag

IEEE

Erscheinungsjahr

2022

Quelle

IEEE Electronic Library Online

Beschreibungen/Notizen

• The traditional fine-grained generation adversarial networks pay less attention to the importance evaluation of word-level information for text-to-image synthesis. Incorrect importance rating of word-level information may skew image generation to less critical direction and affect the generation of key features. In this paper, a novel generative adversarial network based on dynamic word-level update is proposed to solve the above problem by dynamically updating the different importance of each word in the image generation stage. The assignment module with dynamic weights is designed to update text features and image features by communicating the information of text and image. This module enables the importance rating of words to be updated. In addition, the mixed zero-center gradient penalty function and visual loss function are proposed to optimize generative adversarial networks based on dynamic word-level update. The mixed zero-center gradient penalty function allows the generator to generate image with high semantic consistency and ensures the stability of the training process. The visual loss function further improves the visual effect of the generated image by narrowing the difference between the real image and the generated image. Extensive experiments on public benchmark datasets demonstrate that the proposed method outperforms the state-of-the-art methods.