This paper reports the use of the deep learning-based technique to characterize the particle orientation of clay samples. The U-Net model was applied to perform semantic segmentation for identifying individual kaolinite particles, based on the scanning electron microscopic images taken from clay samples subjected to 1-D consolidation. The measurable elongated particles were manually annotated to facilitate the supervised learning. A fivefold cross-validation was used to ensure satisfactory generalization of the deep learning models. Dice loss and weighted cross-entropy were chosen as the loss functions to tackle the issue of imbalanced classification class. The customized weight maps incorporated in the weight cross-entropy were found effective in forcing the deep learning models to learn how to recognize the particle boundaries. With the trained deep learning models, the measurable elongated kaolinite particles were identified from the ~ 1280 patches within ~ 20 min and the particle directional distribution was quantified using the fabric tensor. The characterization results reveal that the kaolinite particles exhibit a tendency to gradually align along the horizontal plane as imposed by the applied vertical stress. In short, the proposed deep learning-based technique is potential to automate the laborious and visually intensive conventional labeling tasks in fabric characterization of clay samples.

本文报告了使用基于深度学习的技术来表征粘土样品的颗粒取向。基于从经受一维固结的粘土样品中获取的扫描电子显微图像，应用 U-Net 模型执行语义分割以识别单个高岭石颗粒。手动注释可测量的细长粒子以促进监督学习。使用五重交叉验证来确保深度学习模型的令人满意的泛化。选择骰子损失和加权交叉熵作为损失函数来解决分类不平衡的问题。发现包含在权重交叉熵中的定制权重图在迫使深度学习模型学习如何识别粒子边界方面是有效的。使用经过训练的深度学习模型，可在约 20 分钟内从约 1280 个斑块中识别出可测量的细长高岭石颗粒，并使用织物张量量化颗粒方向分布。表征结果表明，由于施加的垂直应力，高岭石颗粒表现出沿水平面逐渐排列的趋势。简而言之，所提出的基于深度学习的技术有可能使粘土样品织物表征中费力且视觉密集的传统标记任务自动化。表征结果表明，由于施加的垂直应力，高岭石颗粒表现出沿水平面逐渐排列的趋势。简而言之，所提出的基于深度学习的技术有可能使粘土样品织物表征中费力且视觉密集的传统标记任务自动化。表征结果表明，由于施加的垂直应力，高岭石颗粒表现出沿水平面逐渐排列的趋势。简而言之，所提出的基于深度学习的技术有可能使粘土样品织物表征中费力且视觉密集的传统标记任务自动化。