Scholarlink

<p>The performance of deep learning-based defect detection systems in industrial applications is significantly constrained by the scarcity and class imbalance of annotated defect samples. This paper proposes a Frequency-Structure Collaborative Generative Adversarial Network (FSC-GAN) specifically designed for industrial surface defect data augmentation. Unlike conventional GANs that directly learn global image distributions, FSC-GAN reformulates defect generation as a collaborative distribution modeling problem between texture frequency representations and defect structural representations. The generator integrates a Frequency Consistency Enhancement Module (FCEM) to preserve periodic texture regularity in the Fourier domain and a Geometric Adaptive Structural Encoder (GASE) to capture irregular defect morphologies through adaptive geometric sampling. The discriminator is optimized with Wasserstein distance and gradient penalty to ensure training stability. Experiments on the Alibaba Tianchi tile dataset show that FSC-GAN achieves PSNR of 31.20 dB, SSIM of 0.811, and FID of 21.08, outperforming WGAN-GP, StyleGAN2, and ProGAN. Cross-dataset validation on NEU-DET steel defects confirms generalization (PSNR: 30.15 dB, mAP: 72.31%). Diversity analysis demonstrates effective mode collapse mitigation (LPIPS: 0.24, Recall: 0.81). FSC-GAN improves YOLOv8, RT-DETR, and YOLOv11 by 4.18%, 3.02%, and 5.32% respectively, outperforming few-shot learning and transfer learning while remaining complementary to them. The proposed method offers an effective solution for industrial defect data augmentation with limited training samples.</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

Frequency-structure collaborative generative adversarial network for industrial surface defect synthesis

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月7日

查看全文
  • Xinyi Yu,
  • Xi Chen,
  • Dingran Wang,
  • Xianping Zeng,
  • Jinmin Peng,
  • Bingsan Chen,
  • Yuanfu He

摘要

The performance of deep learning-based defect detection systems in industrial applications is significantly constrained by the scarcity and class imbalance of annotated defect samples. This paper proposes a Frequency-Structure Collaborative Generative Adversarial Network (FSC-GAN) specifically designed for industrial surface defect data augmentation. Unlike conventional GANs that directly learn global image distributions, FSC-GAN reformulates defect generation as a collaborative distribution modeling problem between texture frequency representations and defect structural representations. The generator integrates a Frequency Consistency Enhancement Module (FCEM) to preserve periodic texture regularity in the Fourier domain and a Geometric Adaptive Structural Encoder (GASE) to capture irregular defect morphologies through adaptive geometric sampling. The discriminator is optimized with Wasserstein distance and gradient penalty to ensure training stability. Experiments on the Alibaba Tianchi tile dataset show that FSC-GAN achieves PSNR of 31.20 dB, SSIM of 0.811, and FID of 21.08, outperforming WGAN-GP, StyleGAN2, and ProGAN. Cross-dataset validation on NEU-DET steel defects confirms generalization (PSNR: 30.15 dB, mAP: 72.31%). Diversity analysis demonstrates effective mode collapse mitigation (LPIPS: 0.24, Recall: 0.81). FSC-GAN improves YOLOv8, RT-DETR, and YOLOv11 by 4.18%, 3.02%, and 5.32% respectively, outperforming few-shot learning and transfer learning while remaining complementary to them. The proposed method offers an effective solution for industrial defect data augmentation with limited training samples.