Scholarlink

<p>Wafer microscopy grain segmentation is challenging because weak boundaries, background noise, and texture interference reduce boundary continuity and measurement accuracy. This paper proposes a model-driven framework, termed RMGCCD, which combines entropy-guided contrast enhancement, normalized graph-cut-guided initial segmentation, and corner-constrained narrow-band boundary fusion. The method enhances weak grain rims, generates a stable initial mask, and refines boundary pixels by integrating contour evidence, Canny/gradient edge responses, and Harris corner cues within a narrow band. On single-grain images, RMGCCD achieves a mean IoU of 97.91%, a Dice score of 98.94%, a Boundary-F1 score of 94.08%, and an HD95 of 4.07 pixels. Compared with thresholding, clustering, and U-Net-based baselines, it produces more continuous and geometrically stable boundaries. Multi-grain, ablation, and parameter sensitivity experiments further demonstrate its robustness for metrology-oriented wafer inspection.</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

Enhanced grain segmentation in wafer microscopy using normalized graph-cut-guided boundary fusion

  • 期刊论文

  • 发表日期：

    2026年7月16日

查看全文
  • Pengcheng Ji,
  • Lin Xu,
  • Guo Ye,
  • Zhenzhi He,
  • Xiangning Lu

摘要

Wafer microscopy grain segmentation is challenging because weak boundaries, background noise, and texture interference reduce boundary continuity and measurement accuracy. This paper proposes a model-driven framework, termed RMGCCD, which combines entropy-guided contrast enhancement, normalized graph-cut-guided initial segmentation, and corner-constrained narrow-band boundary fusion. The method enhances weak grain rims, generates a stable initial mask, and refines boundary pixels by integrating contour evidence, Canny/gradient edge responses, and Harris corner cues within a narrow band. On single-grain images, RMGCCD achieves a mean IoU of 97.91%, a Dice score of 98.94%, a Boundary-F1 score of 94.08%, and an HD95 of 4.07 pixels. Compared with thresholding, clustering, and U-Net-based baselines, it produces more continuous and geometrically stable boundaries. Multi-grain, ablation, and parameter sensitivity experiments further demonstrate its robustness for metrology-oriented wafer inspection.