Scholarlink

<p>Deep learning-based visual surface defect detection is widely used for automated quality control in manufacturing. However, its deployment in real-world applications, particularly in scenarios involving new factories, products, processes, and small-batch or multi-batch production, is hindered by significant challenges. These scenarios typically face challenges such as high intra-class variance, high inter-class similarity, defect data scarcity, and severe class imbalance, thereby degrading the performance of detection algorithms. To address these challenges, this study proposes a dual-stage industry image enhancement and generation method (DS-IIEG). In the first stage, to solve the problems of high similarity between defects and backgrounds and large intra class differences, a frequency domain enhancement module was developed to improve defect detail features by enhancing the contrast between positive and negative features in product surface images. This module enhances the detailed features of defects, increases the contrast between defect features and background, and makes defects more recognizable. To address the challenges of data scarcity and class imbalance, the second stage of DS-IIEG incorporates a synthetic data generation module based on a deep convolutional generative adversarial network. This module can generate high-quality and diverse defect images, thereby expanding the dataset and balancing sample distribution. The effectiveness of DS-IIEG was tested on an aluminum alloy part defect dataset collected from a manufacturing site. In addition, we conducted experiments and evaluations using two datasets with single defect features and two datasets with large intra-class differences but similarity between classes. The DS-IIEG method showed significant improvements, with an average F1 score increase of 2.12%, 1.98%, 4.91%, and 4.89%, respectively, across these diverse datasets. These results demonstrate the DS-IIEG effectiveness and feasibility for augmenting small-scale, imbalanced datasets in real-world industrial settings, offering new strategies for intelligent quality inspection and optimizing manufacturing processes. All data and code are available at <a href="https://github.com/xluckywang/DS-IIEG">https://github.com/xluckywang/DS-IIEG</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

DS-IIEG: a dual-stage industrial image enhancement and generation framework for robust surface defect detection in manufacturing with limited data

  • 期刊论文

  • 发表日期：

    2026年1月31日

查看全文
  • Xiaoqiao Wang,
  • Yan Li,
  • Yu Gong,
  • Mingzhou Liu,
  • Kui Luo

摘要

Deep learning-based visual surface defect detection is widely used for automated quality control in manufacturing. However, its deployment in real-world applications, particularly in scenarios involving new factories, products, processes, and small-batch or multi-batch production, is hindered by significant challenges. These scenarios typically face challenges such as high intra-class variance, high inter-class similarity, defect data scarcity, and severe class imbalance, thereby degrading the performance of detection algorithms. To address these challenges, this study proposes a dual-stage industry image enhancement and generation method (DS-IIEG). In the first stage, to solve the problems of high similarity between defects and backgrounds and large intra class differences, a frequency domain enhancement module was developed to improve defect detail features by enhancing the contrast between positive and negative features in product surface images. This module enhances the detailed features of defects, increases the contrast between defect features and background, and makes defects more recognizable. To address the challenges of data scarcity and class imbalance, the second stage of DS-IIEG incorporates a synthetic data generation module based on a deep convolutional generative adversarial network. This module can generate high-quality and diverse defect images, thereby expanding the dataset and balancing sample distribution. The effectiveness of DS-IIEG was tested on an aluminum alloy part defect dataset collected from a manufacturing site. In addition, we conducted experiments and evaluations using two datasets with single defect features and two datasets with large intra-class differences but similarity between classes. The DS-IIEG method showed significant improvements, with an average F1 score increase of 2.12%, 1.98%, 4.91%, and 4.89%, respectively, across these diverse datasets. These results demonstrate the DS-IIEG effectiveness and feasibility for augmenting small-scale, imbalanced datasets in real-world industrial settings, offering new strategies for intelligent quality inspection and optimizing manufacturing processes. All data and code are available at https://github.com/xluckywang/DS-IIEG.