Multi-task deep learning for crack segmentation and quantification in RC structures
Journal
Automation in Construction
Journal Volume
166
Start Page
105599
ISSN
0926-5805
Date Issued
2024-10
Author(s)
Abstract
Crack width is frequently used for damage assessment in reinforced concrete (RC) structures. Recently, deep learning (DL) based approaches have been proposed for crack detection and segmentation. However, most of them focus on extracting the crack regions, followed by width calculation using conventional methods. To address this issue, we propose a multi-task DL model to predict crack segmentation and crack centerline simultaneously. Effects of loss functions are investigated, and state-of-the-art segmentation models are employed as baselines. Results show that the proposed multi-task U-Net model enhances the estimation in crack centerline and outperforms the baselines by more than 2% in width quantification. Moreover, we propose a computer vision (CV) based approach to calculate crack width when camera shooting angle is not perpendicular to target surface. 3D reconstruction and plane fitting are incorporated to correct distortion, and images obtained from non-vertical capturing are used to demonstrate the robustness of the proposed approach.
Subjects
3D reconstruction
Crack segmentation
Crack width quantification
Multi-task deep learning
Plane fitting
SDGs
Publisher
Elsevier BV
Type
journal article