An unsupervised bridge defect identification method based on multiscale masks and global awareness
文献类型：期刊论文
| 作者 | Wang, Yong4; He, Zhenglong3; Cen, Zongxi3; Zeng, Juncheng2; Cheng, Xinyi1 |
| 刊名 | CONSTRUCTION AND BUILDING MATERIALS
 |
| 出版日期 | 2025-10-31 |
| 卷号 | 497页码:143903 |
| 关键词 | Unsupervised deep learning Bridge defects Image segmentation Multiscale masks Global awareness |
| ISSN号 | 0950-0618 |
| DOI | 10.1016/j.conbuildmat.2025.143903 |
| 产权排序 | 1 |
| 文献子类 | Article |
| 英文摘要 | Bridge defects significantly affect structural integrity, necessitating accurate identification and timely maintenance methods for ensuring safety. The existing bridge defect identification models predominantly adopt supervised learning paradigms and rely heavily on abundant labeled data, limiting their generalizability. Although general unsupervised identification models overcome label dependency constraints, they exhibit deficiencies such as insufficient contextual information representations and weak global perception capabilities, yielding poor defect identification performance in complex scenarios and limited applicability to transfer tasks. Therefore, a global strip mask reconstruction network (GSMRNet), the first unsupervised model explicitly developed for bridge defect identification, is proposed herein. Three key specific technical innovations are as follows. First, a novel multiscale random masking strategy enhances contextual encoding and long-range feature correlation processes. Second, a dual-pathway encoder architecture integrating convolutional and Transformer components strengthens local and global information modeling capabilities. Third, a joint loss function combining pixel, structure, and style constraints improves model sensitivity to defect patterns and global contexts. Extensive experiments and analyses are conducted on self-built and publicly available unsupervised defect datasets. Compared with the second-best model, GSMRNet achieves F1 score improvements of 8.72 %, 13.50 %, and 14.85 % on the U-BERD, U-BCD, and U-Crack500 datasets, respectively. GSMRNet accurately identifies bridge defects under unlabeled conditions and effectively captures defect features from different distributions in transfer applications, resulting in robust accuracy and strong generalizability. |
| URL标识 | 查看原文 |
| WOS关键词 | CRACK DETECTION |
| WOS研究方向 | Construction & Building Technology ; Engineering ; Materials Science |
| 语种 | 英语 |
| WOS记录号 | WOS:001592427900005 |
| 出版者 | ELSEVIER SCI LTD |
| 源URL | [http://ir.igsnrr.ac.cn/handle/311030/217435]  |
| 专题 | 资源与环境信息系统国家重点实验室_外文论文 |
| 通讯作者 | Wang, Yong |
| 作者单位 | 1.Fujian Expressway Sci & Technol Innovat Res Inst C, Fuzhou 350001, Peoples R China; 2.Univ Chinese Acad Sci, Coll Resources & Environm, Beijing 100049, Peoples R China; 3.Chinese Acad Sci, Inst Geog Sci & Nat Resources Res, State Key Lab Resources & Environm Informat Syst, Beijing 100101, Peoples R China; 4.China Univ Geosci, Sch Geog & Informat Engn, Wuhan 430070, Peoples R China |
| 推荐引用方式 GB/T 7714 | Wang, Yong,He, Zhenglong,Cen, Zongxi,et al. An unsupervised bridge defect identification method based on multiscale masks and global awareness[J]. CONSTRUCTION AND BUILDING MATERIALS,2025,497:143903. |
| APA | Wang, Yong,He, Zhenglong,Cen, Zongxi,Zeng, Juncheng,&Cheng, Xinyi.(2025).An unsupervised bridge defect identification method based on multiscale masks and global awareness.CONSTRUCTION AND BUILDING MATERIALS,497,143903. |
| MLA | Wang, Yong,et al."An unsupervised bridge defect identification method based on multiscale masks and global awareness".CONSTRUCTION AND BUILDING MATERIALS 497(2025):143903. |
入库方式： OAI收割
来源：地理科学与资源研究所
浏览0
下载0
收藏0
其他版本
除非特别说明，本系统中所有内容都受版权保护，并保留所有权利。