Scholarlink

<p>In the current industrial inspection application, the surface defect data is very limited, extremely hard to acquire, and possesses extremely complex morphology. It is hard for traditional detection models to learn stable feature representations during training due to limited surface-defect data, resulting in insufficient generalization in the wild. This paper proposes a defect generation approach based on Conditional Generative Adversarial Networks (CGAN) with a dual-consistency mechanism that includes structural consistency constraints and defect distribution consistency constraints during generation. This dual-consistency formulation constitutes the central methodological contribution of the proposed surface defect augmentation and detection model. By explicitly separating the structural features of normal samples from the semantic features of defects, the defect injection process is implemented only through local perturbations at semantic locations, while maintaining the structural stability of non-defective regions in both the pixel and perceptual domains. Additionally, statistical feature matching ensures that the generated defects conform to the distribution of real defects in terms of texture and morphology. Experimental results show that the detection model trained on enhanced samples achieves a precision of 0.893 and a recall of 0.969. In contrast, under dual-consistency constraints, the texture feature activation value increases to 0.72, and the structural feature activation value reaches 0.77. The research results indicate that this method has demonstrated strong effectiveness in both defect generation and detection tasks and has application value in improving the intelligence level of industrial visual inspection.</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

Surface defect sample augmentation and detection model based on Generative Adversarial Networks

  • 期刊论文

  • 发表日期：

    2026年6月16日

查看全文
  • Bo Feng,
  • Jiayi Yang,
  • Yi Liu,
  • Yunyun Yao

摘要

In the current industrial inspection application, the surface defect data is very limited, extremely hard to acquire, and possesses extremely complex morphology. It is hard for traditional detection models to learn stable feature representations during training due to limited surface-defect data, resulting in insufficient generalization in the wild. This paper proposes a defect generation approach based on Conditional Generative Adversarial Networks (CGAN) with a dual-consistency mechanism that includes structural consistency constraints and defect distribution consistency constraints during generation. This dual-consistency formulation constitutes the central methodological contribution of the proposed surface defect augmentation and detection model. By explicitly separating the structural features of normal samples from the semantic features of defects, the defect injection process is implemented only through local perturbations at semantic locations, while maintaining the structural stability of non-defective regions in both the pixel and perceptual domains. Additionally, statistical feature matching ensures that the generated defects conform to the distribution of real defects in terms of texture and morphology. Experimental results show that the detection model trained on enhanced samples achieves a precision of 0.893 and a recall of 0.969. In contrast, under dual-consistency constraints, the texture feature activation value increases to 0.72, and the structural feature activation value reaches 0.77. The research results indicate that this method has demonstrated strong effectiveness in both defect generation and detection tasks and has application value in improving the intelligence level of industrial visual inspection.