This paper presents an approach for the integrated process of classification and instance segmentation of leakage‐area and scaling images from shield tunnel linings. For this purpose, the previously established dataset of leakage‐area images by the authors is enlarged by means of adding scaling ones. Afterwards, data augmentation is implemented to enrich the database in the classification dataset, and the augmented classification dataset contains 5776 images. The instance segmentation dataset is subsequently enlarged through original images without any data augmentation, including 1496 images. Then a residual net with 101 layers (i.e., ResNet‐101) is applied to the classification dataset to obtain a model that can identify leakage‐area and scaling images from those of shield tunnel linings. The ResNet‐101 classification model achieves an accuracy of 93.37% in terms of testing classification dataset. Moreover, a mask region‐based convolutional neural network (Mask R‐CNN) is utilized to perform instance segmentation of leakage areas and scaling in the images classified by the ResNet‐101 model. The segmentation results of the Mask R‐CNN model show 96.1% and 95.6% average precision (AP) with intersection over union (IoU) of 0.5 for bounding box and mask predication, respectively. By using the proposed approach, the leakage‐area and scaling defects can be automatically classified and quantified with an overall accuracy of 89.3%, which is quite promising compared to the inherent uncertainty in geotechnical engineering. Image mosaicing is finally applied to provide inspectors with the intuitive observation of the location and distribution information of defects on the tunnel lining.

中文翻译：

---

基于深度学习的泄漏区域分类和实例分割以及盾构隧道衬砌的比例尺图像

本文提出了一种对泄漏区域的分类和实例分割以及盾构隧道衬砌的比例缩放图像的综合过程的方法。为此，作者通过添加缩放比例图像扩大了作者先前建立的泄漏区域图像数据集。之后，实施数据扩充以丰富分类数据集中的数据库，并且扩充的分类数据集中包含5776张图像。随后将实例分割数据集通过原始图像进行放大，而无需进行任何数据增强，包括1496张图像。然后，将具有101层的残差网络（即ResNet-101）应用于分类数据集，以获得一个可以从盾构隧道衬砌中识别泄漏区域和缩放图像的模型。ResNet-101分类模型在测试分类数据集方面的准确性达到93.37％。此外，基于遮罩区域的卷积神经网络（Mask R-CNN）用于对ResNet-101模型分类的图像中的泄漏区域进行实例分割和缩放。Mask R-CNN模型的分割结果显示，边界框和遮罩预测的平均精度（AP）为96.1％，联合相交（IoU）为0.5。通过使用提出的方法，泄漏区域和结垢缺陷可以自动分类和量化，总体精度为89.3％，与岩土工程中固有的不确定性相比，这是很有希望的。