Scholarlink

<p>Enabling Generative Adversarial Networks (GANs) to cover data modes and synthesize new ones remains a significant challenge, involving three key issues: (1) unbalanced competition between the generator and discriminator, especially mode collapse; (2) imbalanced proportions and distributions of rare and common categories, which hinder feature learning; and (3) limited interaction and fusion between different modes despite extensive studies on multimodal feature fusion. To address these problems, this paper proposes Enhancing Diversity Generative Adversarial Networks (EDGAN). Specifically, (1) a diversity discriminator is constructed based on an implicit similarity measure learned by a Siamese Network and incorporated into the traditional GAN framework, providing diversity-oriented feedback in addition to real–fake discrimination; (2) the Wasserstein distance loss with Gradient Penalty is adopted instead of cross-entropy loss to satisfy Lipschitz continuity and improve training stability; and (3) extensive experiments on synthetic and real datasets demonstrate the effectiveness of EDGAN in diverse data generation. Furthermore, the method is applied to creative automotive wheel structure design, where the generated structures satisfy both mechanical and aesthetic requirements, offering a novel approach to automotive exterior design.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

EDGAN: a novel GAN framework incorporating a diversity discriminator for robust data generation

  • 期刊论文

  • 发表日期：

    2026年7月6日

查看全文
  • Aiguo Cheng,
  • Caiyu Ma,
  • Zhicheng He,
  • Hexin Jiang,
  • Eric Li

摘要

Enabling Generative Adversarial Networks (GANs) to cover data modes and synthesize new ones remains a significant challenge, involving three key issues: (1) unbalanced competition between the generator and discriminator, especially mode collapse; (2) imbalanced proportions and distributions of rare and common categories, which hinder feature learning; and (3) limited interaction and fusion between different modes despite extensive studies on multimodal feature fusion. To address these problems, this paper proposes Enhancing Diversity Generative Adversarial Networks (EDGAN). Specifically, (1) a diversity discriminator is constructed based on an implicit similarity measure learned by a Siamese Network and incorporated into the traditional GAN framework, providing diversity-oriented feedback in addition to real–fake discrimination; (2) the Wasserstein distance loss with Gradient Penalty is adopted instead of cross-entropy loss to satisfy Lipschitz continuity and improve training stability; and (3) extensive experiments on synthetic and real datasets demonstrate the effectiveness of EDGAN in diverse data generation. Furthermore, the method is applied to creative automotive wheel structure design, where the generated structures satisfy both mechanical and aesthetic requirements, offering a novel approach to automotive exterior design.