Multi-label bridge damage segmentation with topology supervision and conformal calibration

Reliable automated bridge inspection requires accurate masks, preservation of defect geometry, and uncertainty estimates. This study combines a pretrained ViT-S/16 encoder, a rank-48 residual adapter, progressive decoding, dual-view appearance regularization, boundary and clDice supervision, and label-conditional conformal calibration. DACL10K v2 supplies 19 non-exclusive channels: 13 damage and six component/object classes. Group-disjoint subdivision of its 6935 training images yielded 5202 training, 693 calibration, and 1040 locked internal-test images; 975 official validation images supported model selection and ablation. Native-image evaluation on the internal test yielded Macro-IoU 0.758, Macro-Dice 0.861, Micro-IoU 0.817, and Boundary F1 0.792. ECE was 0.016 and the Brier score was 0.094. Five-seed validation ablation yielded Macro-IoU 0.803 ± 0.003 for the full model. Seventeen controlled perturbations identified haze and occlusion as major failure conditions; stronger compound corruptions reduced positive-label conformal coverage to 0.793–0.849. Matched five-seed retraining gave Macro-IoU 0.757 ± 0.003 for Full and 0.766 ± 0.004 for Mask2Former/Swin-T, showing no overlap superiority of the proposed backbone. These results support a joint internal assessment of overlap, geometry, and uncertainty. External-domain transfer and field robustness remain unvalidated.

Authors

Institutions

Publication Details

Journal
Scientific Reports
Published
2026-11-21
DOI
https://doi.org/10.1038/s41598-026-81581-z
Primary Topic
Infrastructure Maintenance and Monitoring
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Multi-label bridge damage segmentation with topology supervision and conformal calibration

Yusufbay Yusupov, Sodikova Nargiza, Komilov Kamol, Seyedjavad Rafiei et al.
Scientific Reports
Infrastructure Maintenance and Monitoring
article

Multi-label bridge damage segmentation with topology supervision and conformal calibration

Yusufbay Yusupov, Sodikova Nargiza, Komilov Kamol, Seyedjavad Rafiei, Sardor Kenjayev, Doston Sayfiddinov, Ghasem Dabouri Farimani
article en

Abstract

Reliable automated bridge inspection requires accurate masks, preservation of defect geometry, and uncertainty estimates. This study combines a pretrained ViT-S/16 encoder, a rank-48 residual adapter, progressive decoding, dual-view appearance regularization, boundary and clDice supervision, and label-conditional conformal calibration. DACL10K v2 supplies 19 non-exclusive channels: 13 damage and six component/object classes. Group-disjoint subdivision of its 6935 training images yielded 5202 training, 693 calibration, and 1040 locked internal-test images; 975 official validation images supported model selection and ablation. Native-image evaluation on the internal test yielded Macro-IoU 0.758, Macro-Dice 0.861, Micro-IoU 0.817, and Boundary F1 0.792. ECE was 0.016 and the Brier score was 0.094. Five-seed validation ablation yielded Macro-IoU 0.803 ± 0.003 for the full model. Seventeen controlled perturbations identified haze and occlusion as major failure conditions; stronger compound corruptions reduced positive-label conformal coverage to 0.793–0.849. Matched five-seed retraining gave Macro-IoU 0.757 ± 0.003 for Full and 0.766 ± 0.004 for Mask2Former/Swin-T, showing no overlap superiority of the proposed backbone. These results support a joint internal assessment of overlap, geometry, and uncertainty. External-domain transfer and field robustness remain unvalidated.

Scientific Reports
Bukhara State University (UZ), Tashkent Institute of Architecture and Civil Engineering (UZ), Kimyo International University in Tashkent (UZ), Karakalpak State University (UZ), University of Tashkent for Applied Sciences (UZ), Buxoro davlat pedagogika instituti (UZ), ZARMED University (UZ), University of Houston (US), Iran University of Science and Technology (IR)
Infrastructure Maintenance and Monitoring
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon