Geometry-informed deep learning-based structural component segmentation of post-earthquake buildings. (1st April 2023)

• Record Type: Journal Article
• Authors: Wang, Yu; Jing, Xin; Chen, Wenli; Li, Hui; Xu, Yang; Zhang, Qiangqiang
• Is Part Of: Mechanical systems and signal processing. Volume 188(2023)
• Publication Date: 2023-04-01
• Subjects: Geometry-informed deep learning -- Geometric consistency loss -- Enhanced UNet -- Structural component segmentation -- Post-earthquake assessment; Structural dynamics -- Periodicals; Vibration -- Periodicals; Constructions -- Dynamique -- Périodiques; Vibration -- Périodiques; Structural dynamics; Vibration; Periodicals; 621
• DOI: 10.1016/j.ymssp.2022.110028
• Languages: English
• ISSNs: 0888-3270

Abstract

Abstract: Traditional manual visual inspections have demonstrated certain shortcomings in post-earthquake assessment of urban buildings, such as being time-consuming and laborious. In contrast, computer vision (CV) and unmanned aerial vehicle (UAV) approaches have revealed competitive potentials in the fields of automatic data acquisition, data processing, and autonomous decision-making. In UAV images, structural components of post-earthquake buildings often present different scales, which are affected by different local damage. Therefore, acquiring the feature information of structural components has precisely been significant for refined damage assessment of post-earthquake buildings. This study proposes a geometry-informed deep learning-based structural component segmentation of post-earthquake buildings. An Enhanced UNet model is established with a new synthetical loss function containing the geometric consistency (GC) term. Given an edge closure of a connected domain for homogeneous structural components, the GC term comprises split line loss and area loss to adapt to the circumference and area constraints of each component region. The Enhanced UNet network is designed to improve the extraction capability of high-level features, and it includes six encoder stages (superior to five in the original version), of which the bottom four stages have many convolution layers, and five corresponding decoders. The investigated synthetic QuakeCity dataset includes 4, 809 images with a resolution of 1, 920 × 1, 080 pixels. Training and test results reveal that compared to the original UNet, the proposed method achieves a more stable training process and higher test accuracy for structural component segmentation. The proposed method can achieve a mIoU of 97.97 %, which is 1.29 % higher than that of the original UNet. In addition, misrecognition of inner voids inside structural components is addressed, which further validates the optimization efficiency of the proposed geometric constraints. Ablation experiments are conducted to confirm the effectiveness of the proposed GC loss and Enhanced UNet network. The proposed method shows good generalization ability in robustness tests in complex real-world scenarios under various disturbances, including abnormal exposure and rain lines in various intensities.