Scholarlink

<p>Industrial defect detection faces fundamental challenges due to scarce defective samples and inadequate training diversity, particularly limiting deep learning model performance in real-world ceramic tile manufacturing scenarios. To address the constraints of conventional data augmentation techniques, which provide insufficient diversity and limited morphological variations, we propose Tile-DCGAN, a comprehensively enhanced Deep Convolutional Generative Adversarial Network for high-fidelity defect image synthesis. Our methodology incorporates three principal innovations: (1) Wasserstein distance with gradient penalty supersedes traditional loss formulations to eliminate gradient vanishing and mode collapse phenomena, (2) a multi-scale generator architecture integrating Feature Pyramid Networks with Frequency-Spatial Attention mechanisms to capture complex defect characteristics across diverse spatial resolutions, and (3) deformable convolution (DCNv4) within the discriminator to enable adaptive sampling of irregular defect geometries and spatial distributions. Comprehensive evaluation on the Tianchi tile defect dataset demonstrates superior image quality metrics compared to existing generative approaches. Furthermore, downstream detection experiments across multiple state-of-the-art frameworks show significant improvement in performance, validating the practical effectiveness of our synthetic data augmentation strategy in industrial quality control applications.</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

Tile-DCGAN: a tile defect image generation method

  • 期刊论文

  • 发表日期：

    2026年2月17日

查看全文
  • Xiaoyu Wang,
  • Meishun Wu,
  • Yong Yang,
  • Chunling Wang

摘要

Industrial defect detection faces fundamental challenges due to scarce defective samples and inadequate training diversity, particularly limiting deep learning model performance in real-world ceramic tile manufacturing scenarios. To address the constraints of conventional data augmentation techniques, which provide insufficient diversity and limited morphological variations, we propose Tile-DCGAN, a comprehensively enhanced Deep Convolutional Generative Adversarial Network for high-fidelity defect image synthesis. Our methodology incorporates three principal innovations: (1) Wasserstein distance with gradient penalty supersedes traditional loss formulations to eliminate gradient vanishing and mode collapse phenomena, (2) a multi-scale generator architecture integrating Feature Pyramid Networks with Frequency-Spatial Attention mechanisms to capture complex defect characteristics across diverse spatial resolutions, and (3) deformable convolution (DCNv4) within the discriminator to enable adaptive sampling of irregular defect geometries and spatial distributions. Comprehensive evaluation on the Tianchi tile defect dataset demonstrates superior image quality metrics compared to existing generative approaches. Furthermore, downstream detection experiments across multiple state-of-the-art frameworks show significant improvement in performance, validating the practical effectiveness of our synthetic data augmentation strategy in industrial quality control applications.