Scholarlink

<p>Few-shot industrial defect detection is hindered by the scarcity and diversity of anomalous samples, whereas standard geometric or mixing-based augmentations often fail to introduce realistic, label-preserving appearance variations. We developed a prompt-free diffusion-based data augmentation framework that synthesizes pseudo-defect images via masked img2img inpainting guided by geometric pattern masks and ControlNet constraints. Binary masks with simple geometric patterns specify regions to be regenerated, enabling local appearance diversification while preserving global structure; two ControlNet branches (Lineart and Reference-Only) further stabilize geometry and transfer style from the source image. We evaluated the method on two industrial benchmarks, MVTec AD and VisA, under class-balanced few-shot settings (5–25 images per class) using two backbones (EfficientNetV2-M and ViT-B/16), and compared against representative baselines including standard transforms, MixUp, and CutMix. On MVTec AD, our augmentation consistently improved average accuracy across all shot regimes, reaching 0.918 (EfficientNetV2-M, 20-shot) and 0.870 (ViT-B/16, 20-shot), compared to 0.869 and 0.778 without augmentation, respectively. On VisA, the method achieved considerable gains in low-shot regimes and was consistently effective for ViT-B/16, improving the average accuracy from 0.583 to 0.696 at 5-shot and from 0.739 to 0.828 at 25-shot. Ablation studies validated the contributions of control guidance and mask design, and further experiments demonstrated that combining the proposed augmentation with conventional transforms (e.g., horizontal flip) provides additional improvements.</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

Few-Shot Industrial Defect Detection via Diffusion-Based Data Augmentation with Geometric Pattern Masks

  • 期刊论文

  • 发表日期：

    2026年6月1日

查看全文
  • Masaya Oirase,
  • Eisuke Kita

摘要

Few-shot industrial defect detection is hindered by the scarcity and diversity of anomalous samples, whereas standard geometric or mixing-based augmentations often fail to introduce realistic, label-preserving appearance variations. We developed a prompt-free diffusion-based data augmentation framework that synthesizes pseudo-defect images via masked img2img inpainting guided by geometric pattern masks and ControlNet constraints. Binary masks with simple geometric patterns specify regions to be regenerated, enabling local appearance diversification while preserving global structure; two ControlNet branches (Lineart and Reference-Only) further stabilize geometry and transfer style from the source image. We evaluated the method on two industrial benchmarks, MVTec AD and VisA, under class-balanced few-shot settings (5–25 images per class) using two backbones (EfficientNetV2-M and ViT-B/16), and compared against representative baselines including standard transforms, MixUp, and CutMix. On MVTec AD, our augmentation consistently improved average accuracy across all shot regimes, reaching 0.918 (EfficientNetV2-M, 20-shot) and 0.870 (ViT-B/16, 20-shot), compared to 0.869 and 0.778 without augmentation, respectively. On VisA, the method achieved considerable gains in low-shot regimes and was consistently effective for ViT-B/16, improving the average accuracy from 0.583 to 0.696 at 5-shot and from 0.739 to 0.828 at 25-shot. Ablation studies validated the contributions of control guidance and mask design, and further experiments demonstrated that combining the proposed augmentation with conventional transforms (e.g., horizontal flip) provides additional improvements.