Scholarlink

Accurate defect detection is critical for industrial quality control systems, yet remains constrained by severe class imbalance in defect distributions. While generative adversarial networks (GANs) have shown potential for defect sample generation, conventional GAN architectures struggle with minority-class defect synthesis due to their inability to capture subtle anomaly characteristics and disentangle defect-specific features. This paper presents DefGAN-Im, an enhanced adversarial framework built upon FastGAN—a lightweight GAN variant employing self-supervised learning—with two innovations: First, we introduce a conditional variational autoencoder (CVAE) pretraining module that establishes label-aware latent space constraints through variational inference, enabling explicit disentanglement of defect-specific feature representations. Second, we develop a multi-task discriminator architecture that unifies authenticity discrimination and fine-grained defect classification within a joint optimization space, therefore forcing the network to preserve discriminative defect patterns during generation, mitigate mode collapse, and accelerate convergence. Experiments demonstrate DefGAN-Im’s superiority over state-of-the-art methods in generating realistic, diverse defect samples, and achieving significant improvements in downstream detection tasks under imbalanced data conditions.

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

DefGAN-Im: Adversarial Industrial Defect Synthesis with Conditional Feature Disentanglement on Imbalanced Datasets

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Xiaofei Nan,
  • Yiming Zhu,
  • Jinyu Fan,
  • Wenyang Li,
  • Yuzhuo Wang,
  • Xiaoheng Jiang

摘要

Accurate defect detection is critical for industrial quality control systems, yet remains constrained by severe class imbalance in defect distributions. While generative adversarial networks (GANs) have shown potential for defect sample generation, conventional GAN architectures struggle with minority-class defect synthesis due to their inability to capture subtle anomaly characteristics and disentangle defect-specific features. This paper presents DefGAN-Im, an enhanced adversarial framework built upon FastGAN—a lightweight GAN variant employing self-supervised learning—with two innovations: First, we introduce a conditional variational autoencoder (CVAE) pretraining module that establishes label-aware latent space constraints through variational inference, enabling explicit disentanglement of defect-specific feature representations. Second, we develop a multi-task discriminator architecture that unifies authenticity discrimination and fine-grained defect classification within a joint optimization space, therefore forcing the network to preserve discriminative defect patterns during generation, mitigate mode collapse, and accelerate convergence. Experiments demonstrate DefGAN-Im’s superiority over state-of-the-art methods in generating realistic, diverse defect samples, and achieving significant improvements in downstream detection tasks under imbalanced data conditions.