Scholarlink

<p>Industrial defect detection in complex environments faces difficulties due to background clutter and varied defect morphologies, which complicate precise segmentation. To address this, we propose the Curvature Pyramid Vision Graph Neural Network (CPViG), a novel multi-scale graph segmentation framework driven by discrete Ricci curvature. CPViG integrates curvature-guided graph construction with K-Nearest Neighbor (KNN)-based feature affinity and multi-scale graph convolutions, enabling holistic structural modeling across multiple receptive fields. It performs unstructured convolution of defect features via an improved Curvature Flow-based graph clustering mechanism for accurate defect region localization. This architecture reduces background interference, thereby improving the model’s stability in complex settings. Evaluations across multiple benchmark industrial defect datasets confirm that CPViG achieves significant improvements in both the precision of defect delineation and resilience to complex backgrounds over conventional CNN-based approaches.</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

Ricci curvature enhance the vision graph neural network for defect detection

  • 期刊论文

  • 发表日期：

    2026年2月10日

查看全文
  • Yili Chen,
  • Huiling Huang,
  • Hanfei Lin,
  • YueMing Hu,
  • Zhenting Yan,
  • Xingzhao Hua,
  • Jun Han

摘要

Industrial defect detection in complex environments faces difficulties due to background clutter and varied defect morphologies, which complicate precise segmentation. To address this, we propose the Curvature Pyramid Vision Graph Neural Network (CPViG), a novel multi-scale graph segmentation framework driven by discrete Ricci curvature. CPViG integrates curvature-guided graph construction with K-Nearest Neighbor (KNN)-based feature affinity and multi-scale graph convolutions, enabling holistic structural modeling across multiple receptive fields. It performs unstructured convolution of defect features via an improved Curvature Flow-based graph clustering mechanism for accurate defect region localization. This architecture reduces background interference, thereby improving the model’s stability in complex settings. Evaluations across multiple benchmark industrial defect datasets confirm that CPViG achieves significant improvements in both the precision of defect delineation and resilience to complex backgrounds over conventional CNN-based approaches.