Automatic bridge crack detections: A lightweight segmentation model for UAV inspection
4 affiliations
- 1Key Laboratory of Opto-Electronic Technology and Intelligent Control, Ministry of Education, Lanzhou Jiaotong University, Lanzhou, China
- 2School of Electronics and Information Engineering, Lanzhou Jiaotong University, Lanzhou, China
- 3National and Provincial Joint Engineering Laboratory of Road & Bridge Disaster Prevention and Control, Lanzhou Jiaotong University, Lanzhou, China
- 4School of Materials Science and Engineering, Southeast University, Nanjing, China
Abstract
Bridge cracks are among the primary factors affecting structural safety and durability. Intelligent and precise detection and quantification of cracks are important in improving bridge operation and maintenance efficiency and safety assessment. Traditional manual detection methods are inefficient and highly subjective, while existing deep learning methods still have shortcomings in terms of computational overhead and detailed modelling. Therefore, this study proposes a high-precision and lightweight crack semantic segmentation network—Feather SFASwinUnet. The model introduces two new modules: Scale Fusion Attention (SFA) and Feather ConvBlock (FCB). SFA effectively fuses global semantic information and local detailed features through a dual-branch structure, which improves spatial structure modelling capabilities. Meanwhile, FCB compensates for feature distortion with extremely low overhead in skip connections, which enhances the recovery capability of fine-grained cracks. Experimental results show that, on the public dataset SDNET2018, the proposed model achieves an mIoU value of 81.25% and an mDice value of 88.57%. It significantly outperforms typical methods such as Unet, DeepLabV3+ and SwinUnet while maintaining low computational load. Furthermore, the YOLO11-BD detection algorithm is combined to achieve crack detection and segmentation under complex backgrounds, which effectively improves robustness. Lastly, the orthogonal skeleton method is used to quantitatively analyse the segmentation results. As a result, the measurement errors of crack width and length are controlled within 5%, which provides reliable technical support for bridge structural health assessment and maintenance decisions. The Feather SFASwinUnet proposed in this study provides an efficient and scalable technical approach for automatic crack detection and quantitative analysis of bridges under UAV inspection. In addition, it has high engineering application potential.
Related
- FDUNet-LRSA: efficient frequency-adaptive segmentation for automated crop disease inspection
- Multistage AI Crop Inspection and Segmentation Framework (MACISF)
- TEFAR: a lightweight, configurable framework for semi-automatic artefact component classification in EEG and TMS-EEG
- Segmentation and morphological analysis of recanalized coiled aneurysms: Towards an automatic tool for clipping planning in microsurgery
- Semi-automatic segmentation of pelvic bone tumors: Usability testing
BibTeX
@article{dong2026automatic,
title = {{Automatic bridge crack detections: A lightweight segmentation model for UAV inspection}},
author = {Xuwei Dong and Zhibing Yu and Jiashuo Yuan and Jinpeng Dai},
journal = {PLOS One},
year = {2026},
month = sep,
doi = {10.1371/journal.pone.0358865},
url = {https://doi.org/10.1371/journal.pone.0358865},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Automatic bridge crack detections: A lightweight segmentation model for UAV inspection
AU - Xuwei Dong
AU - Zhibing Yu
AU - Jiashuo Yuan
AU - Jinpeng Dai
PY - 2026
DA - 2026/09/24
JO - PLOS One
DO - 10.1371/journal.pone.0358865
UR - https://doi.org/10.1371/journal.pone.0358865
AB - Bridge cracks are among the primary factors affecting structural safety and durability. Intelligent and precise detection and quantification of cracks are important in improving bridge operation and maintenance efficiency and safety assessment. Traditional manual detection methods are inefficient and highly subjective, while existing deep learning methods still have shortcomings in terms of computational overhead and detailed modelling. Therefore, this study proposes a high-precision and lightweight crack semantic segmentation network—Feather SFASwinUnet. The model introduces two new modules: Scale Fusion Attention (SFA) and Feather ConvBlock (FCB). SFA effectively fuses global semantic information and local detailed features through a dual-branch structure, which improves spatial structure modelling capabilities. Meanwhile, FCB compensates for feature distortion with extremely low overhead in skip connections, which enhances the recovery capability of fine-grained cracks. Experimental results show that, on the public dataset SDNET2018, the proposed model achieves an mIoU value of 81.25% and an mDice value of 88.57%. It significantly outperforms typical methods such as Unet, DeepLabV3+ and SwinUnet while maintaining low computational load. Furthermore, the YOLO11-BD detection algorithm is combined to achieve crack detection and segmentation under complex backgrounds, which effectively improves robustness. Lastly, the orthogonal skeleton method is used to quantitatively analyse the segmentation results. As a result, the measurement errors of crack width and length are controlled within 5%, which provides reliable technical support for bridge structural health assessment and maintenance decisions. The Feather SFASwinUnet proposed in this study provides an efficient and scalable technical approach for automatic crack detection and quantitative analysis of bridges under UAV inspection. In addition, it has high engineering application potential.
SN - 1932-6203
ER -