In this study, a novel approach of processing μCT images to create digital material twins is presented. A deep convolutional neural network (DCNN) was implemented and used to segment μCT images of two different types of reinforcement (2D glass and 3D carbon). The DCNN successfully segmented the images based on multi-scale features extracted using data-driven convolutional filters. The network was trained using scanned μCT images, along with images extracted from computer-generated virtual models of the reinforcements. One of the convolutional layers of the trained network was utilized to extract features to be used in creating the machine learning-based model. The extracted features and the raw gray-scale data was used to train a supervised k-nearest neighbor (k-NN) model for pixel-wise classification. The performance of both approaches was evaluated by comparing the results with manually segmented images. The trained deep neural network was able to provide faster and superior predictions of different features of the reinforcements as compared to a conventional machine learning approach.

在这项研究中，提出了一种处理μCT图像以创建数字材料双胞胎的新颖方法。实施了深度卷积神经网络（DCNN），并将其用于分割两种不同类型的增强材料（2D玻璃和3D碳）的μCT图像。DCNN基于使用数据驱动卷积滤波器提取的多尺度特征成功地对图像进行了分割。该网络使用扫描的μCT图像以及从计算机生成的钢筋虚拟模型中提取的图像进行训练。训练网络的卷积层之一被用来提取特征，以用于创建基于机器学习的模型。提取的特征和原始灰度数据用于训练有监督的k最近邻（kNN）模型用于像素分类。通过将结果与手动分割的图像进行比较，评估了这两种方法的性能。与传统的机器学习方法相比，训练有素的深度神经网络能够对钢筋的不同特征提供更快，更好的预测。