当前位置: X-MOL 学术Comput. Aided Civ. Infrastruct. Eng. › 论文详情
Our official English website, www.x-mol.net, welcomes your feedback! (Note: you will need to create a separate account there.)
Automatic determination of 3D particle morphology from multiview images using uncertainty‐evaluated deep learning
Computer-Aided Civil and Infrastructure Engineering ( IF 8.5 ) Pub Date : 2025-01-14 , DOI: 10.1111/mice.13421
Hongchen Liu, Huaizhi Su, Brian Sheil

Particle morphology is a crucial factor influencing the mechanical properties of granular materials particularly in infrastructure construction processes where accurate shape descriptors are essential. Accurately measuring three‐dimensional (3D) morphology has significant theoretical and practical value for exploring the multiscale mechanical properties of civil engineering materials. This study proposes a novel approach using multiview (two‐dimensional [2D]) particle images to efficiently predict 3D morphology, making real‐time aggregate quality analysis feasible. A 3D convolutional neural network (CNN) model is developed, which combines Monte Carlo dropout and attention mechanisms to achieve uncertainty‐evaluated predictions of 3D morphology. The model incorporates a convolutional block attention module, involving a two‐stage attention mechanism with channel attention and spatial attention, to further optimize feature representation and enhance the effectiveness of the attention mechanism. A new dataset comprising 18,000 images of 300 natural gravel and 300 blasted rock fragment particles is used for model training. The prediction accuracy and uncertainty of the proposed model are benchmarked against a range of alternative models including 2D CNN, 3D CNN, and 2D CNN with attention, in particular, to the influence of the number of input multiview particle images on the performance of the models for predicting various morphological parameters is explored. The results indicate that the proposed 3D CNN model with the attention mechanism achieves high prediction accuracy with an error of less than 10%. Whilst it exhibits initially greater uncertainty compared to other models due to its increased complexity, the model shows significant improvement in both accuracy and uncertainty as the number of training images is increased. Finally, residual challenges associated with the prediction of more complex particle angles and irregular shapes are also discussed.

中文翻译:


使用不确定性评估深度学习从多视图图像中自动确定 3D 颗粒形态



颗粒形态是影响颗粒材料机械性能的关键因素,尤其是在基础设施建设过程中,准确的形状描述符是必不可少的。精确测量三维 (3D) 形貌对于探索土木工程材料的多尺度力学性能具有重要的理论和实践价值。本研究提出了一种使用多视图(二维 [2D])颗粒图像的新方法来有效预测 3D 形态,使实时聚集体质量分析成为可能。开发了一个 3D 卷积神经网络 (CNN) 模型,该模型结合了蒙特卡洛辍学和注意力机制,以实现 3D 形态的不确定性评估预测。该模型包含一个卷积块注意力模块,涉及通道注意力和空间注意力的两阶段注意力机制,以进一步优化特征表示并增强注意力机制的有效性。一个新的数据集包含 300 个天然砾石和 300 个爆破岩石碎片颗粒的 18000 张图像,用于模型训练。所提出的模型的预测准确性和不确定性与包括 2D CNN、3D CNN 和 2D CNN 在内的一系列替代模型进行了基准测试,特别注意输入多视图粒子图像的数量对模型预测各种形态参数性能的影响。结果表明,所提出的具有注意力机制的三维卷积神经网络模型实现了较高的预测精度,误差小于10%。 虽然由于复杂性增加,与其他模型相比,它最初表现出更大的不确定性,但随着训练图像数量的增加,该模型在准确性和不确定性方面都显示出显着提高。最后,还讨论了与预测更复杂的粒子角度和不规则形状相关的残余挑战。
更新日期:2025-01-14
down
wechat
bug