Scholarlink

<p>In fabric manufacturing, surface defects are unavoidable, and their accurate detection and documentation are considered essential for quality assessment. Although deep learning based inspection methods have shown potential for reducing labor costs and improving efficiency, their performance is limited by the highly homogeneous textures of plain fabrics, where defect regions exhibit subtle variations with weak contrast. Conventional detection frameworks often fail to learn sufficiently discriminative features, resulting in degraded accuracy. To address these challenges, a lightweight feature enhancement framework is proposed, focusing on local statistical modeling, structural anomaly reinforcement, and industrial deployment. First, a texture-aware local statistical perception mechanism is introduced, in which neighborhood-level statistical references characterize stable background textures and serve as anomaly baselines, allowing defect-induced deviations to be effectively distinguished. Second, a structural anomaly driven nonlinear enhancement is applied, amplifying weak signals associated with edge disruption and texture continuity breakdown, improving the detection of low-contrast and structurally complex defects. Ablation experiments demonstrate the effectiveness of the proposed modules. Specifically, the Average Precision at an Intersection over Union threshold of 0.5 improved from 71.9% to 74.0%, while the recall rate reached 91.2%. Despite the modest gain, this approach demonstrates robust performance across varying lighting conditions, effectively balancing detection precision and recall. The method has also been successfully deployed in real-world industrial fabric inspection systems, enabling real-time defect detection, localization, and reporting under practical operating conditions while maintaining high inference efficiency. In addition, valuable insights for product quality control are offered by the proposed machine vision based fabric defect detection framework, and significant practical potential for intelligent manufacturing is exhibited.</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

Structure-aware feature enhancement for real-time fabric defect detection in industrial inspection

  • 期刊论文

  • 发表日期：

    2026年7月11日

查看全文
  • Shangyan Zhang,
  • Haohao Xie,
  • JunNan Lu,
  • ZhiKe Ji,
  • Jian Zhou,
  • Jun Xiang,
  • Ruru Pan

摘要

In fabric manufacturing, surface defects are unavoidable, and their accurate detection and documentation are considered essential for quality assessment. Although deep learning based inspection methods have shown potential for reducing labor costs and improving efficiency, their performance is limited by the highly homogeneous textures of plain fabrics, where defect regions exhibit subtle variations with weak contrast. Conventional detection frameworks often fail to learn sufficiently discriminative features, resulting in degraded accuracy. To address these challenges, a lightweight feature enhancement framework is proposed, focusing on local statistical modeling, structural anomaly reinforcement, and industrial deployment. First, a texture-aware local statistical perception mechanism is introduced, in which neighborhood-level statistical references characterize stable background textures and serve as anomaly baselines, allowing defect-induced deviations to be effectively distinguished. Second, a structural anomaly driven nonlinear enhancement is applied, amplifying weak signals associated with edge disruption and texture continuity breakdown, improving the detection of low-contrast and structurally complex defects. Ablation experiments demonstrate the effectiveness of the proposed modules. Specifically, the Average Precision at an Intersection over Union threshold of 0.5 improved from 71.9% to 74.0%, while the recall rate reached 91.2%. Despite the modest gain, this approach demonstrates robust performance across varying lighting conditions, effectively balancing detection precision and recall. The method has also been successfully deployed in real-world industrial fabric inspection systems, enabling real-time defect detection, localization, and reporting under practical operating conditions while maintaining high inference efficiency. In addition, valuable insights for product quality control are offered by the proposed machine vision based fabric defect detection framework, and significant practical potential for intelligent manufacturing is exhibited.