Scholarlink

In the manufacturing industry, defect classification is crucial but is hampered by challenges from imbalanced data, which often leads to model overfitting when data are scarce. One common solution is to use synthetic data from generative models; however, these models typically produce results that are structurally inconsistent. Addressing these concerns, this paper introduces a novel null embedding optimization technique that generates latent representations closely resembling the original images, significantly enhancing the fidelity of the generated images. This method ensures that synthetic images are not only visually similar to the original images but also subjected to a more extensive and diverse augmentation process, increasing the variability within the dataset. Consequently, this approach effectively doubles the usable dataset size and notably increases the accuracy of classification models by up to 11%. For those interested, the dataset and source code are accessible at https://github.com/ugiugi0823/DISN .

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

Deterministic Synthesis of Defect Images Using Null Optimization

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Hyunwook Jo,
  • Marcellino Sahadewa,
  • William Gazali,
  • In Kyu Park

摘要

In the manufacturing industry, defect classification is crucial but is hampered by challenges from imbalanced data, which often leads to model overfitting when data are scarce. One common solution is to use synthetic data from generative models; however, these models typically produce results that are structurally inconsistent. Addressing these concerns, this paper introduces a novel null embedding optimization technique that generates latent representations closely resembling the original images, significantly enhancing the fidelity of the generated images. This method ensures that synthetic images are not only visually similar to the original images but also subjected to a more extensive and diverse augmentation process, increasing the variability within the dataset. Consequently, this approach effectively doubles the usable dataset size and notably increases the accuracy of classification models by up to 11%. For those interested, the dataset and source code are accessible at https://github.com/ugiugi0823/DISN .