Scholarlink

<p>Industrial quartz is vital to high-end manufacturing, where accurate surface inspection helps reduce supply chain loss. We present QiddNet, an unsupervised reconstruction framework that addresses multi-scale defects and generalization to unseen anomalies. Instead of proposing a new anomaly detection paradigm, QiddNet integrates a multi-scale squeeze–excite RepViTBlock (MS-SERVB) for dynamic multi-scale defect feature capture, an adaptive residual mask self-attention (ARMSA) mechanism for focusing on defect regions, and a dynamic feature fusion-up-sampling (DF-Up) module for high-fidelity image reconstruction. Our constructed industrial quartz defect dataset (IQDD) demonstrates QiddNet’s superiority, achieving an image-level AUROC of 98.6%, a pixel-level AUROC of 94.9%, and a per-region overlap of 80.6%. Cross-domain experiments further confirmed its excellent generalization ability, with average accuracies of 99.1% and 98.7% for I-AUROC and 94.4% and 98.4% for P-AUROC on the public datasets MVTec AD and MPDD, respectively. These results highlight QiddNet’s potential to significantly improve industrial quartz defect detection accuracy and localization precision. Code and data are available at <a href="https://zenodo.org/records/16956951">https://zenodo.org/records/16956951</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

QiddNet: enhancing industrial quartz defect detection via unsupervised reconstruction

  • 期刊论文

  • 发表日期：

    2026年1月20日

查看全文
  • Haipeng Pan,
  • Haoyu Wang,
  • Minming Gu

摘要

Industrial quartz is vital to high-end manufacturing, where accurate surface inspection helps reduce supply chain loss. We present QiddNet, an unsupervised reconstruction framework that addresses multi-scale defects and generalization to unseen anomalies. Instead of proposing a new anomaly detection paradigm, QiddNet integrates a multi-scale squeeze–excite RepViTBlock (MS-SERVB) for dynamic multi-scale defect feature capture, an adaptive residual mask self-attention (ARMSA) mechanism for focusing on defect regions, and a dynamic feature fusion-up-sampling (DF-Up) module for high-fidelity image reconstruction. Our constructed industrial quartz defect dataset (IQDD) demonstrates QiddNet’s superiority, achieving an image-level AUROC of 98.6%, a pixel-level AUROC of 94.9%, and a per-region overlap of 80.6%. Cross-domain experiments further confirmed its excellent generalization ability, with average accuracies of 99.1% and 98.7% for I-AUROC and 94.4% and 98.4% for P-AUROC on the public datasets MVTec AD and MPDD, respectively. These results highlight QiddNet’s potential to significantly improve industrial quartz defect detection accuracy and localization precision. Code and data are available at https://zenodo.org/records/16956951.