Scholarlink

<p>The fragmentation distribution of a blasted rock mass is a critical parameter for understanding its geomechanical response and optimizing excavation designs. However, current computer vision methods struggle with two fundamental problems: the prohibitive cost of pixel-level data annotation, and severe fragment adhesion (under-segmentation) in densely packed heap blocks, which fundamentally distorts the quantitative block-size distribution. To address these limitations, this study proposes a decoupled semantic-topological framework that bridges deep learning with macroscopic geomechanical statistics. Specifically, the core contributions of the proposed method are threefold. First, a self-supervised semantic segmentation architecture (MAE-UPerNet) is introduced, leveraging unannotated datasets to extract robust domain-invariant features and drastically reduce annotation reliance. Second, a composite loss function is incorporated into the UPerNet head to mitigate extreme sample imbalance between dominant boulders and sparse boundary gaps. Third, to explicitly resolve the fatal challenge of adhered blocks, a marker-controlled watershed algorithm is developed, overcoming the geometric flaws of standard bounding-box paradigms. For large-scale macroscopic heap analysis, UAV image stitching using SIFT and RANSAC algorithms is additionally employed. Experimental results demonstrate that the proposed method not only achieves high pixel-wise accuracy (85.36% mIoU and 90.51% mAcc) but also significantly improves contour-level separation, verified via Boundary IoU. Most crucially, the algorithm-derived Particle Size Distribution (PSD) curve achieves a highly robust statistical fit with the manual ground truth (R<sup>2</sup> = 0.9314), proving its exceptional engineering reliability. This study provides an accurate, automated, and generalized technical support framework for intelligent blast quality assessment in open-pit mining.</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

Self-supervised segmentation of large-scale blasted heap block from UAV image: addressing block adhesion and imbalanced samples

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月21日

查看全文
  • Binjian Rao,
  • Song Jiang,
  • Caiwu Lu,
  • Qinghua Gu,
  • Zhixiang Cui,
  • Guowei Jiang,
  • Guanghe Li

摘要

The fragmentation distribution of a blasted rock mass is a critical parameter for understanding its geomechanical response and optimizing excavation designs. However, current computer vision methods struggle with two fundamental problems: the prohibitive cost of pixel-level data annotation, and severe fragment adhesion (under-segmentation) in densely packed heap blocks, which fundamentally distorts the quantitative block-size distribution. To address these limitations, this study proposes a decoupled semantic-topological framework that bridges deep learning with macroscopic geomechanical statistics. Specifically, the core contributions of the proposed method are threefold. First, a self-supervised semantic segmentation architecture (MAE-UPerNet) is introduced, leveraging unannotated datasets to extract robust domain-invariant features and drastically reduce annotation reliance. Second, a composite loss function is incorporated into the UPerNet head to mitigate extreme sample imbalance between dominant boulders and sparse boundary gaps. Third, to explicitly resolve the fatal challenge of adhered blocks, a marker-controlled watershed algorithm is developed, overcoming the geometric flaws of standard bounding-box paradigms. For large-scale macroscopic heap analysis, UAV image stitching using SIFT and RANSAC algorithms is additionally employed. Experimental results demonstrate that the proposed method not only achieves high pixel-wise accuracy (85.36% mIoU and 90.51% mAcc) but also significantly improves contour-level separation, verified via Boundary IoU. Most crucially, the algorithm-derived Particle Size Distribution (PSD) curve achieves a highly robust statistical fit with the manual ground truth (R2 = 0.9314), proving its exceptional engineering reliability. This study provides an accurate, automated, and generalized technical support framework for intelligent blast quality assessment in open-pit mining.