Related Experiment Video
Updated: Jul 16, 2026
06:45
Automated Joint Space Detection Improves Bone Segmentation Accuracy
Published on: November 28, 2025
A Structure-Aware Deep Learning Framework for Automated Bridge Inspection Integrating SegFormer-Based Structural
Sushama De Silva1, Pang-Jo Chun1
1Institute of Engineering Innovation, School of Engineering, The University of Tokyo, Tokyo 113-8656, Japan.
Sensors (Basel, Switzerland)
|July 15, 2026
Summary
This study introduces a structure-aware deep learning framework for automated bridge inspection, accurately associating damage like cracks and corrosion with specific structural members. This AI-assisted system enhances maintenance decision-making for aging bridge infrastructure.
Area of Science:
- Civil Engineering
- Artificial Intelligence
- Computer Vision
- Structural Health Monitoring
Background:
- Aging bridge infrastructure necessitates automated and reliable condition assessment systems due to limited inspection resources.
- Existing deep learning methods often fail to associate detected damage with specific structural members, hindering practical maintenance planning.
- There is an urgent need for advanced AI solutions to improve the efficiency and accuracy of bridge inspections.
Purpose of the Study:
- To propose and evaluate a novel structure-aware deep learning framework for automated bridge inspection.
- To integrate structural member segmentation, damage detection, and damage-to-member association within a unified pipeline.
- To provide structured, actionable information for bridge maintenance decision-making.
Main Methods:
- Developed a unified deep learning pipeline incorporating SegFormer for structural member segmentation (main girder, deck slab, abutment) and YOLOv8s for damage detection (crack, corrosion).
- Utilized the Segment Anything Model (SAM) for boundary refinement in training data preparation.
- Implemented a region-based spatial assignment strategy to associate detected damage with segmented structural members.
Main Results:
- Achieved a mean Intersection over Union (mIoU) of 0.851 for structural member segmentation.
- Attained a mean Average Precision (mAP50) of 0.445 for detecting cracks and corrosion.
- Demonstrated high accuracy in associating damage with members, with fully correct and partially correct accuracies of 62.0% and 87.0%, respectively.
Conclusions:
- The proposed structure-aware framework successfully associates detected damage with specific structural members, offering practical utility for bridge inspection.
- The system provides structured outputs (e.g., 'crack on main girder') to support maintenance assessment and infrastructure monitoring.
- This AI-assisted approach shows significant potential as a foundational tool for enhancing bridge maintenance strategies.
