Scholarlink

We propose a generalized data augmentation framework for industrial defect detection that leverages conditionally controlled diffusion models in combination with geometric-pattern-based masks. Unlike conventional image transformation and augmentation methods, our approach enables the diverse synthesis of rare anomalous data while maintaining the structural characteristics of original samples, even under severe data scarcity. By generalizing previous mask-based diffusion model techniques beyond narrow domains such as leaf disease classification, the presented method flexibly generates realistic anomalous samples for a broad range of industrial scenarios. Experiments on the MVTec AD dataset demonstrate that our framework consistently outperforms standard augmentation approaches across varying data set sizes—for example, improving classification accuracy from 76.5% to 79.5% in the five-shot setting and from 86.9% to 91.8% with twenty-shot training. Notably, the use of text prompts is not required, reducing the need for prompt engineering, and additional performance gains are realized when our method is combined with conventional augmentations. These results highlight both the effectiveness and practical deployability of diffusion-model-based augmentation with geometric masking for robust industrial anomaly detection.

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

Data Augmentation Using Diffusion Models with Geometric Pattern Masks for Industrial Defect Detection

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Masaya Oirase,
  • Eisuke Kita

摘要

We propose a generalized data augmentation framework for industrial defect detection that leverages conditionally controlled diffusion models in combination with geometric-pattern-based masks. Unlike conventional image transformation and augmentation methods, our approach enables the diverse synthesis of rare anomalous data while maintaining the structural characteristics of original samples, even under severe data scarcity. By generalizing previous mask-based diffusion model techniques beyond narrow domains such as leaf disease classification, the presented method flexibly generates realistic anomalous samples for a broad range of industrial scenarios. Experiments on the MVTec AD dataset demonstrate that our framework consistently outperforms standard augmentation approaches across varying data set sizes—for example, improving classification accuracy from 76.5% to 79.5% in the five-shot setting and from 86.9% to 91.8% with twenty-shot training. Notably, the use of text prompts is not required, reducing the need for prompt engineering, and additional performance gains are realized when our method is combined with conventional augmentations. These results highlight both the effectiveness and practical deployability of diffusion-model-based augmentation with geometric masking for robust industrial anomaly detection.