Chengjia Han, Yiqing Dong, Maggie Y. Gao, Liwei Dong, Yaowen Yang
{"title":"Unsupervised anomaly segmentation model for rail damage based on image-inpainting and cold diffusion","authors":"Chengjia Han, Yiqing Dong, Maggie Y. Gao, Liwei Dong, Yaowen Yang","doi":"10.1016/j.autcon.2025.106342","DOIUrl":null,"url":null,"abstract":"<div><div>Ensuring structural health of rail tracks is critical for safe train operations. While deep learning-based vision models are widely used for rail damage detection, supervised methods suffer from limited generalization due to scarce and diverse annotated data. Unsupervised models often experience missed detections and false positives when handling complex and variable rail background textures, as well as rail damage with significant intra-class variability. To address these limitations, this paper proposes an unsupervised pixel-level rail damage segmentation model based on a cold diffusion framework, called InpRailDiffusion. It introduces inpainting-based noise and uses a Mamba-enhanced, time-conditioned U-Net for progressive noise removal. Damage segmentation is achieved by analyzing pixel-wise differences between generated and original images with adaptive thresholding. A multi-scale masking strategy fuses reconstruction features at various spatial resolutions, reducing false positives and missed detections. Evaluated on RSDDs-I and RSDDs-II, InpRailDiffusion outperformed state-of-the-art baselines with MIoU/F1-Scores of 0.864/0.844 and 0.845/0.814, respectively.</div></div>","PeriodicalId":8660,"journal":{"name":"Automation in Construction","volume":"177 ","pages":"Article 106342"},"PeriodicalIF":9.6000,"publicationDate":"2025-06-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Automation in Construction","FirstCategoryId":"5","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S0926580525003826","RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"CONSTRUCTION & BUILDING TECHNOLOGY","Score":null,"Total":0}
引用次数: 0
Abstract
Ensuring structural health of rail tracks is critical for safe train operations. While deep learning-based vision models are widely used for rail damage detection, supervised methods suffer from limited generalization due to scarce and diverse annotated data. Unsupervised models often experience missed detections and false positives when handling complex and variable rail background textures, as well as rail damage with significant intra-class variability. To address these limitations, this paper proposes an unsupervised pixel-level rail damage segmentation model based on a cold diffusion framework, called InpRailDiffusion. It introduces inpainting-based noise and uses a Mamba-enhanced, time-conditioned U-Net for progressive noise removal. Damage segmentation is achieved by analyzing pixel-wise differences between generated and original images with adaptive thresholding. A multi-scale masking strategy fuses reconstruction features at various spatial resolutions, reducing false positives and missed detections. Evaluated on RSDDs-I and RSDDs-II, InpRailDiffusion outperformed state-of-the-art baselines with MIoU/F1-Scores of 0.864/0.844 and 0.845/0.814, respectively.
期刊介绍:
Automation in Construction is an international journal that focuses on publishing original research papers related to the use of Information Technologies in various aspects of the construction industry. The journal covers topics such as design, engineering, construction technologies, and the maintenance and management of constructed facilities.
The scope of Automation in Construction is extensive and covers all stages of the construction life cycle. This includes initial planning and design, construction of the facility, operation and maintenance, as well as the eventual dismantling and recycling of buildings and engineering structures.