Scholarlink

<p>Current saliency-based defect detection methods show promise in industrial settings, but the unpredictability of defects in industrial production environments complicates dataset creation, which in turn hampers model performance. While existing data expansion approaches using generative models show potential, they typically require large quantities of defect or defect-free samples, which can be impractical in industrial settings. To address these challenges, we introduce DefFiller, a novel mask-conditioned generation method that achieves precise pixel-level control even with limited training data by leveraging layout-to-image diffusion prior. Our method generates high-quality defect samples paired with corresponding mask conditions that can be directly utilized for detection model training. Additionally, we develop an evaluation framework to assess the quality of generated samples and their impact on detection performance. Experimental results on two industrial defect datasets demonstrate that DefFiller produces high-quality defect images that accurately match the provided mask conditions, significantly enhancing the performance of saliency-based defect detection models trained on the augmented dataset. The code is available at: <a href="https://github.com/CC-T/DefFiller">https://github.com/CC-T/DefFiller</a>.</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

DefFiller: mask-conditioned generation with diffusion prior for saliency-based defect detection

  • 期刊论文

  • 发表日期：

    2025年7月16日

查看全文
  • Yichun Tai,
  • Zhenzhen Huang,
  • Tao Peng,
  • Zhijiang Zhang

摘要

Current saliency-based defect detection methods show promise in industrial settings, but the unpredictability of defects in industrial production environments complicates dataset creation, which in turn hampers model performance. While existing data expansion approaches using generative models show potential, they typically require large quantities of defect or defect-free samples, which can be impractical in industrial settings. To address these challenges, we introduce DefFiller, a novel mask-conditioned generation method that achieves precise pixel-level control even with limited training data by leveraging layout-to-image diffusion prior. Our method generates high-quality defect samples paired with corresponding mask conditions that can be directly utilized for detection model training. Additionally, we develop an evaluation framework to assess the quality of generated samples and their impact on detection performance. Experimental results on two industrial defect datasets demonstrate that DefFiller produces high-quality defect images that accurately match the provided mask conditions, significantly enhancing the performance of saliency-based defect detection models trained on the augmented dataset. The code is available at: https://github.com/CC-T/DefFiller.