基于深度学习的数字钻孔图像裂隙、岩脉与岩溶识别算法研究

Research on recognition algorithm of fractures, veins and kavst in digital borehole images based on deep learning

摘要

【目的】岩体结构面参数的精确获取对深部地下工程的稳定性评价至关重要。现有数字钻孔图像的人工识别方法存在主观性强、效率低等问题，而基于图像处理的自动识别方法在应对结构类型多样化与识别鲁棒性不足等挑战时，准确率仍有限。为此，提出一种基于深度学习的裂隙、岩脉与岩溶识别算法，旨在实现结构面的高效识别和高精度参数提取。【方法】针对裂隙、岩脉和岩溶这三类结构的精细分割需求，构建了多目标语义分割标签体系，设计了融合频谱门控模块、动态通道注意力机制、注意力门控单元与深度监督机制的SpectDA-ResU-Net分割模型，显著增强了复杂结构的识别准确性。此外，提出了一种基于分割掩膜的三维几何参数自动提取方法。【结果】在2 148张钻孔图像的增强数据集上进行的消融试验显示，所提模型集成所有模块后，F1-score达到94.46%(提升4.65%),mIoU为89.59%(提升8.69%)。在与U-Net模型的对比试验中，mIoU提升了15.77%,各项指标也显著优于现有主流分割网络。贵州岩溶区的实例分析表明，产状自动提取误差控制在4%以内。【结论】研究表明，所提算法在提高数字钻孔图像结构面识别精度与参数提取效率方面具有显著优势。

Abstract

[Objective] The precise acquisition of rock mass structural plane parameters is crucial for the stability evaluation of deep underground engineering. Existing manual recognition method for digital borehole images suffer from subjectivity and low efficiency, while image processing-based automatic recognition method still face challenges in accuracy due to the diversity of structural types and insufficient recognition robustness. Therefore, this paper proposes a deep learning-based algorithm for the identification of fractures, veins, and karst, aiming to achieve efficient recognition of structural planes and high-precision parameter extraction. [Methods] In response to the fine segmentation requirements for fractures, veins, and karst, a multi-target semantic segmentation label system is constructed. The SpectDA-ResU-Net segmentation model is designed, integrating spectral gating modules, dynamic channel attention mechanisms, attention gating units, and deep supervision mechanisms, significantly enhancing the accuracy of complex structure recognition. Additionally, an automatic 3D geometric parameter extraction method based on segmentation masks is proposed. [Results] Ablation experiments on an enhanced dataset of 2 148 borehole images show that after integrating all modules, the proposed model achieves an F1-score of 94.46%(an improvement of 4.65%) and an mIoU of 89.59%(an improvement of 8.69%). In comparison with the U-Net model, the mIoU improves by 15.77%, and other evaluation metrics are significantly better than those of existing mainstream segmentation networks. Case studies in the Guizhou karst region show that the automatic extraction error of structural attitudes is controlled within 4%. [Conclusion] The research demonstrates that the proposed algorithm has significant advantages in improving the accuracy of digital borehole image structural plane recognition and the efficiency of parameter extraction.

关键词

数字钻孔图像 / 深度学习 / 结构面 / 产状 / 智能识别 / 影响因素 / 多目标语义分割 / 贵州岩溶区

Key words

digital borehole images / deep learning / structural plane / orientation / intelligent recognition / influencing factors / multi-target semantic segmentation / Guizhou karst region

基金资助

国家重点研发计划重点专项项目(2022YFC3705000)

国家重点研发计划重点专项项目(2022YFC3705001)

贵州省地热水和矿泉水顶尖专家团队(黔科合人才CXTD[2025]003)