Scholarlink

<p>Rock mass discontinuities strongly control rock mass strength, deformability, and fluid flow, making accurate discontinuity mapping essential for geological engineering, mining, and civil infrastructure. Conventional field mapping is labor-intensive and subjective, while current CNN-based discontinuity extraction still faces practical constraints, including (i) performance degradation under variable illumination, (ii) inefficient training/inference for high-resolution outcrop images, and (iii) fragmented discontinuity traces that hinder engineering interpretation. Here, rock mass images are categorized into small-scale and large-scale datasets, and three scale-aware strategies are proposed. First, for small-scale datasets, an MGV-based clustering and augmentation scheme mitigates illumination-induced performance variability, with particularly stable improvement for low-illumination images. Second, for large-scale images, a splitting-and-merging strategy enables efficient high-resolution processing without loss of segmentation performance. Third, a Hough-transform-based correction with subsequent post-processing reconnects fragmented traces while suppressing redundant detections and abnormal branching, thereby improving trace connectivity. The proposed framework systematically presents a complete workflow for applying CNN-based rock mass discontinuity extraction under field conditions, and the effectiveness of each step was experimentally validated.</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

Automated CNN-based rock discontinuity mapping with a refined correction method addressing image lighting variability, resolution constraints, and connectivity issues

  • 期刊论文

  • 发表日期：

    2026年5月23日

查看全文
  • Sang Seob Kim,
  • Gyung Won Lee,
  • Tae Hyeon Kim,
  • Melvin B. Diaz,
  • Tae Sup Yun,
  • Kwang Yeom Kim

摘要

Rock mass discontinuities strongly control rock mass strength, deformability, and fluid flow, making accurate discontinuity mapping essential for geological engineering, mining, and civil infrastructure. Conventional field mapping is labor-intensive and subjective, while current CNN-based discontinuity extraction still faces practical constraints, including (i) performance degradation under variable illumination, (ii) inefficient training/inference for high-resolution outcrop images, and (iii) fragmented discontinuity traces that hinder engineering interpretation. Here, rock mass images are categorized into small-scale and large-scale datasets, and three scale-aware strategies are proposed. First, for small-scale datasets, an MGV-based clustering and augmentation scheme mitigates illumination-induced performance variability, with particularly stable improvement for low-illumination images. Second, for large-scale images, a splitting-and-merging strategy enables efficient high-resolution processing without loss of segmentation performance. Third, a Hough-transform-based correction with subsequent post-processing reconnects fragmented traces while suppressing redundant detections and abnormal branching, thereby improving trace connectivity. The proposed framework systematically presents a complete workflow for applying CNN-based rock mass discontinuity extraction under field conditions, and the effectiveness of each step was experimentally validated.