Multiclass Concrete Surface Defect Segmentation and Quantification via an Improved U-Net with VGG16-CBAM
Rapid detection and quantification of concrete surface defects are essential for structural condition assessment. However, given the disparate geometric characteristics of fine cracks, exposed rebar, and large-area spalling, many existing deep learning–based methods are limited to recognizing only a single defect category. This study proposes an automated framework for segmenting and measuring cracks, spalling, and exposed rebar in high-resolution concrete images. An improved U-Net is developed by using a VGG16 encoder and a convolutional block attention module to enhance defect feature discrimination. A sliding-window inference and coordinate-based stitching strategy is adopted to obtain full-resolution segmentation masks. Based on calibration, crack length and width are extracted from skeletonized masks, while spalling and exposed rebar areas are calculated by connected-component analysis. Experimental results show that the proposed framework provides effective pixel-level segmentation and engineering-scale geometric measurements for concrete surface defects, with relative errors of crack width, crack length, spalling area, and exposed rebar area all maintained within 7%. The proposed method provides technical support for bridge inspectors and postdisaster assessment teams in the condition evaluation of damaged components and engineering decision-making.
Authors
- Sulan Li (ORCID: https://orcid.org/0000-0002-5156-6421)
- Shibin Lin (ORCID: https://orcid.org/0000-0002-8410-5584)
- Peipei Yuan (ORCID: https://orcid.org/0000-0002-7773-5780)
- Peng Wan (ORCID: https://orcid.org/0009-0005-1666-7541)
- Su Yan (ORCID: https://orcid.org/0009-0006-9769-6551)
Institutions
- Jianghan University (CN)
- China Railway Group (China) (CN)
Publication Details
- Journal
- Journal of Performance of Constructed Facilities
- Published
- 2026-11-21
- DOI
- https://doi.org/10.1061/jpcfev.cfeng-5681
- Primary Topic
- Infrastructure Maintenance and Monitoring
- Type
- article
- Field-Weighted Citation Impact
- 0.00