Bridge corrosion is usually recorded as a plain fact in an inspection report: rust is there, or it is not. A research team at Saitama University in Japan wants inspection software to say more than that. Their two-level AI framework first marks every pixel of visible corrosion in a photo of a steel bridge. It then sorts those pixels into four visual patterns that inspectors already recognise: Uniform, Crevice, Underfilm and Other Localized corrosion.
The work comes from PhD student Sal Saad Al Deen Taher and Associate Professor Ji Dang of the university’s Department of Civil and Environmental Engineering. It was published online on 5 August 2026 in the journal Computer-Aided Civil and Infrastructure Engineering, and Tech Xplore reported it on 15 September. “Conventional AI-based corrosion segmentation can tell us where corrosion is visible, but inspectors often need more information than a simple yes-or-no corrosion map,” Taher said.
Below we explain how the two levels work, what the four bridge corrosion categories mean and how the models scored on 1,960 and 1,738 inspection images. We also show where those scores are strong or weak, set the study against the ageing bridge stock in Japan and the US, compare it with other published approaches, and list what bridge owners can do with the idea today. For the wider landscape of models behind this kind of research, see our AI models and tools hub.
Table of contents
- What Saitama University Built for Bridge Corrosion Inspection
- The Four Bridge Corrosion Categories in the Framework
- Inside the Bridge Corrosion Datasets
- How the Models Scored on Bridge Corrosion
- What the Scores Mean for Bridge Corrosion Inspectors
- Why Bridge Corrosion Is an Ageing-Infrastructure Problem
- How the Framework Compares With Other Bridge Corrosion AI Research
- Limits of the Two-Level Bridge Corrosion Framework
- What Bridge Owners Can Do With Bridge Corrosion AI Now
- Frequently Asked Questions About Bridge Corrosion AI
- References
What Saitama University Built for Bridge Corrosion Inspection
The paper is titled “A two-level semantic segmentation framework for visual corrosion pattern assessment in steel bridges”. Semantic segmentation means the model labels each pixel of an image with a class. It does not draw a box around an object or give the whole photo a single label. That pixel-level output is what makes a bridge corrosion map possible: an inspector can see an overlay showing exactly which parts of a girder, bearing or bracing member look rusted.
Level 1 finds where the corrosion is
Level 1 is a binary task. Each pixel is either corrosion or background, and the model’s job is to capture the presence and spatial extent of visible corrosion. This is the layer where most existing bridge corrosion tools stop. It answers the first question an inspector asks about a member: how much of it is affected, and where.
Level 2 says what kind of corrosion it is
Level 2 takes the image and assigns each corroded pixel to one of four inspection-oriented visual categories. Together with the background class, that makes a five-class problem. The output is still a map, but it is now coded by pattern. A crevice at a plate connection and a broad patch of surface rust no longer look identical on screen.
Why a yes-or-no map is not enough
Two members can show the same rusted area and still call for very different responses. A wide film of even surface rust usually means a coating has worn away. A narrow line of rust bleeding from a joint can point to trapped water inside a connection. “Our framework adds another layer by distinguishing visually different corrosion patterns at the pixel level,” Taher said.
The researchers describe the combined output as complementary information about three things. It shows where corrosion occurs, what visual pattern it exhibits and how accurately its boundaries are represented. That third point matters, and we return to it when we look at the scores.
The Four Bridge Corrosion Categories in the Framework
The study groups visible bridge corrosion into four categories. The names come from the paper. The descriptions below are general corrosion-engineering explanations of what each pattern usually looks like on a steel bridge. They are not definitions quoted from the study, which has not published its full annotation guide in the public summaries.
Uniform corrosion
Uniform corrosion spreads fairly evenly across an exposed steel surface. On a bridge it typically appears as broad rusted areas on webs, flanges or the steel under a deck where a coating has failed over a wide area. It removes metal slowly and predictably. That predictability is why it is often the easiest bridge corrosion pattern to plan maintenance around.
Crevice corrosion
Crevice corrosion develops in narrow gaps where water and salts are trapped and fresh oxygen cannot circulate. Typical places are the joints between connected plates, the ground under bolt heads and the edges of splice plates. On steel bridges these crevices sit at connections, which is exactly where lost section can matter most for load paths. It can also advance faster than the open surface around it.
Underfilm corrosion
Underfilm corrosion grows beneath a paint or coating film. It often shows as blistering, bubbling or thread-like lines under a surface that still looks mostly intact. For an inspector it is an early warning that a protective system is failing before bare rust spreads. That makes it a timing signal for recoating decisions rather than a strength problem on its own.
Other Localized corrosion
The fourth class collects localized attack that does not fit the first three. In corrosion engineering, localized attack includes concentrated spots such as pitting. The study treats everything in this group as one visual class, so the model does not separate pitting from other localized mechanisms. That keeps the labelling task manageable, at the cost of some detail.
| Category | Typical appearance | Where it tends to occur | Why it matters |
|---|---|---|---|
| Uniform | Broad, even rust over a wide area | Webs, flanges, deck underside | Slow, predictable metal loss; plan recoating |
| Crevice | Rust lines or staining at gaps | Plate joints, bolt heads, splice edges | Sits at connections; can advance quickly |
| Underfilm | Blisters or threads under paint | Coated members that still look intact | Early sign a coating system is failing |
| Other Localized | Concentrated spots of attack | Anywhere water or salt collects | Small area, but may hide deeper loss |
Inside the Bridge Corrosion Datasets
The images came from Japanese steel bridge inspection reports. That is an important detail. These are field photographs taken by inspectors during routine work, with the uneven lighting, awkward angles and cluttered backgrounds that real bridges produce. They are not clean laboratory panels photographed under controlled light.
Two datasets for two levels
The team built a binary corrosion dataset of 1,960 images for Level 1 and a categorised dataset of 1,738 images for Level 2. The categorised set is 222 images smaller, or 88.7% of the binary set (1,738 ÷ 1,960). A smaller set for the harder task is common, because every corroded pixel in it needs a pattern label as well as a yes-or-no label.
How much agreement the labels reached
Pixel labels for bridge corrosion are partly a judgement call, especially at the fuzzy edge where rust fades into sound paint. The researchers therefore measured how reproducible their annotations were. Overall agreement was 92.81%, and Cohen’s kappa was 0.892.
Kappa corrects raw agreement for the agreement two annotators would reach by chance. On the widely used Landis and Koch scale, values between 0.81 and 1.00 count as almost perfect agreement, so 0.892 sits comfortably in the top band. It is a reassuring number, because a model can never be more consistent than the labels it learns from. We looked at why that matters in our piece on hidden annotation errors in object detection datasets.
What the public summary does not say
The public summaries do not give the split between training, validation and test images, the number of bridges photographed, or how many pixels fall into each of the four classes. Class balance matters for a five-class task. If Uniform corrosion dominates the photos, a model can score well on average while struggling with a rarer pattern such as Underfilm.
| Item | Reported value | What it tells you |
|---|---|---|
| Image source | Japanese steel bridge inspection reports | Real field conditions, one country |
| Level 1 dataset | 1,960 images, binary labels | Where corrosion is visible |
| Level 2 dataset | 1,738 images, four categories | 88.7% of the binary set |
| Overall agreement | 92.81% | Labels were largely reproducible |
| Cohen’s kappa | 0.892 | Almost perfect band after chance correction |
| Train/test split, class balance | Not in public summaries | Needed to judge rare-class accuracy |
How the Models Scored on Bridge Corrosion
The researchers tested both convolutional neural network models and transformer-based models. Both are standard computer vision architectures for segmentation. Convolutional networks build up features from small local patches of pixels. Transformers use attention to relate distant parts of an image, which can help with long, thin or scattered shapes like rust streaks.
SegFormer-B2 leads Level 1
For finding bridge corrosion, the best overall model was SegFormer-B2 trained with boundary-aware supervision, meaning the training gives extra weight to getting edges right. It reached a mean intersection over union (mIoU) of 84.30%, a corrosion-class IoU of 73.39% and an F1 score of 84.65%. SegFormer is a transformer design first published by NVIDIA researchers in 2021, offered in sizes from B0 (smallest) to B5 (largest).
Mask2Former-Small draws the cleanest edges
A different model won on outlines. Mask2Former-Small recorded the best Boundary-F1 score, 69.10%. Mask2Former is a universal segmentation architecture published by Meta AI researchers. Its edge result supports the researchers’ point that boundary accuracy is a separate quality to measure, not something a high average score guarantees.
SegFormer-B5 leads Level 2
For classifying corrosion patterns, the largest SegFormer variant did best. SegFormer-B5 reached a five-class mIoU of 60.45% and a mean F1 score of 67.81%. That is a respectable result for telling apart visual categories that can overlap in the same photo, but it is clearly lower than the Level 1 scores.
The chart below plots the six best scores exactly as reported, so the bar lengths equal the percentages.
How to read IoU, F1 and Boundary-F1
Intersection over union compares the pixels a model marks with the pixels the annotators marked. It divides the overlap by the combined area, so 100% means a perfect match. Mean IoU averages that across classes, including background, which is why the Level 1 mIoU (84.30%) is higher than the corrosion-only IoU (73.39%). Background pixels are easier to get right.
The F1 score balances precision (how much of what the model flags is really corrosion) and recall (how much real corrosion it finds). Boundary-F1 applies the same idea only to pixels near the edges of each region. A model can find the right patches of bridge corrosion and still draw their outlines loosely, and Boundary-F1 is the number that exposes it.
| Metric | Level | Best model | Score | Plain meaning |
|---|---|---|---|---|
| mIoU | 1 | SegFormer-B2 | 84.30% | Average overlap, corrosion and background |
| Corrosion IoU | 1 | SegFormer-B2 | 73.39% | Overlap on the corrosion class alone |
| F1 | 1 | SegFormer-B2 | 84.65% | Balance of false alarms and misses |
| Boundary-F1 | 1 | Mask2Former-Small | 69.10% | How tightly outlines follow the rust |
| Five-class mIoU | 2 | SegFormer-B5 | 60.45% | Average overlap across patterns and background |
| Mean F1 | 2 | SegFormer-B5 | 67.81% | Balance of errors per pattern, averaged |
What the Scores Mean for Bridge Corrosion Inspectors
Headline accuracy figures are easy to over-read. The useful questions for an inspection team are how big the gap is between finding and classifying, where the models are weakest, and what the output can and cannot replace.
A 23.85-point drop from finding to classifying
The best Level 2 mIoU (60.45%) is 23.85 points below the best Level 1 mIoU (84.30%). The mean F1 falls by 16.84 points, from 84.65% to 67.81%. The two tasks are not directly comparable, because one has two classes and the other has five. Still, the drop shows that naming the pattern is much harder than spotting the rust, which is also true for human inspectors.
Edges are the weak spot
The best Boundary-F1, 69.10%, is 15.55 points below the best Level 1 F1 of 84.65%. Those numbers come from different models and different metrics, so the comparison is indicative only. It still matters in practice. Any estimate of corroded area, and any trend drawn from comparing photos across inspection cycles, depends on where the outline of bridge corrosion is drawn.
A triage layer, not a condition rating
The framework describes what corrosion looks like. It does not measure remaining steel thickness or section loss, and the public summaries do not map its output to a formal condition state. So it is best read as a triage and documentation layer. It can flag a crevice pattern at a joint for closer attention, but an engineer still decides what that means for the structure.
| Question an inspector asks | Answered by the framework? | Notes |
|---|---|---|
| Is there visible corrosion, and where? | Yes, Level 1 | Best F1 84.65% |
| What visual pattern is it? | Yes, Level 2 | Best mean F1 67.81% |
| How precisely are edges drawn? | Measured | Best Boundary-F1 69.10% |
| How much steel has been lost? | No | Needs thickness or physical measurement |
| Is the member still safe? | No | Structural judgement by an engineer |
Why Bridge Corrosion Is an Ageing-Infrastructure Problem
The research is aimed at a practical bottleneck. Steel bridges built in the second half of the twentieth century are reaching the age at which coatings fail and bridge corrosion accelerates, and inspection rules require regular, documented visual checks of every one of them.
Japan’s bridge stock and the five-year rule
Japan has about 730,000 road bridges, and 520,000 of them, or 70%, sit on municipal roads, according to the Ministry of Land, Infrastructure, Transport and Tourism (MLIT). The ministry’s road maintenance page projects that within ten years 48% of all bridges will be 50 years or older. After a tunnel ceiling collapse in 2012, a 2014 ministerial ordinance obliged road administrators to carry out a close visual inspection once every five years.
The arithmetic of that rule explains the interest in automation. Spreading 730,000 bridges across five years means about 146,000 bridge inspections a year on average (730,000 ÷ 5). Each one produces photographs that someone must review, describe and file.
The US inventory in 2025
The US National Bridge Inventory counted 624,193 bridges in 2025. Of those, 272,779 were rated good, 309,729 fair and 41,685 poor. That works out to 43.7% good, 49.6% fair and 6.7% poor. Routine US inspections have a baseline interval of up to 24 months, so the inventory generates hundreds of thousands of inspections every year.
The chart below uses those 2025 counts divided by the 624,193 total, so each bar shows a share of the inventory.
Why photographs pile up
Condition ratings are only as good as the evidence behind them. In both countries, photos are the cheapest evidence to collect and the slowest to interpret at scale. A tool that consistently labels bridge corrosion patterns across thousands of archived images could help owners compare one inspection cycle with the next, spot members where a pattern is changing, and give less experienced inspectors a consistent second opinion.
| Factor | Japan | United States |
|---|---|---|
| Bridge stock | About 730,000 road bridges | 624,193 bridges (2025) |
| Inspection rule | Close visual inspection every 5 years since 2014 | Routine inspections, baseline up to 24 months |
| Average annual load | About 146,000 bridges a year | Hundreds of thousands a year |
| Age or condition signal | MLIT projects 48% will be 50+ years old within a decade | 6.7% rated poor in 2025 |
How the Framework Compares With Other Bridge Corrosion AI Research
The Saitama study joins a busy field. Its contribution is not the idea of segmenting rust, which others have done, but the second layer of visual pattern classes plus a separate measure of boundary quality.
Detection and condition rating
A 2024 study in the journal Infrastructures used Mask R-CNN and YOLOv8 to segment steel bridge corrosion and support condition rating. Another group, writing in the Journal of Civil Structural Health Monitoring, applied transformer-based segmentation to stitched views of bridges. That work was evaluated on a public dataset built by Virginia Tech researchers from Virginia Department of Transportation inspection reports and annotated to AASHTO condition-state guidance.
Element-level corrosion intensity
A 2026 paper on quantifying element-level corrosion intensity took a different route. It first identifies which structural element a pixel belongs to, then characterises the defect, and fuses the two. Its dual Mask2Former design reported an average F1 of 94.5% for structural element identification. That answers a question the Saitama framework leaves open: which member of the bridge the corrosion is on.
Probability of detection for cracks
An August 2026 preprint evaluated AI-based visual crack detection in steel bridges using probability of detection, a reliability measure borrowed from non-destructive testing. It is about cracks rather than bridge corrosion, but the method matters. Asset owners understand probability-of-detection curves, and they may ask for the same kind of evidence from corrosion tools.
| Approach | What it outputs | Main strength | What it leaves open |
|---|---|---|---|
| Two-level framework (Saitama, 2026) | Corrosion map plus four pattern classes | Pattern detail and boundary scoring | Member identity, section loss |
| Mask R-CNN and YOLOv8 rating (2024) | Corrosion segments and condition support | Links detection to rating | Visual pattern types |
| Stitched-view transformer (2025) | Corrosion levels across stitched images | Whole-member views | Pattern classes |
| Element-level late fusion (2026) | Corrosion intensity per structural element | Knows which member is affected | Pattern classes |
| Crack probability of detection (2026) | Detection reliability curves | Speaks the language of NDT | Corrosion, not covered |
Where the two-level approach fits
These methods look complementary rather than competing. A future inspection pipeline could identify the structural element, map bridge corrosion on it, classify the pattern and report a reliability figure for each step. The Saitama framework supplies the pattern step, and its separate boundary metric is a useful habit for the others to copy.
Limits of the Two-Level Bridge Corrosion Framework
The researchers are open about what has not been tested yet. Their stated next steps say a lot about how far the method is from routine field use.
One country’s photographs
All images came from Japanese inspection reports. Coating systems, climate, de-icing salt use and photo practices vary by country and by owner. The team says the framework needs testing on independent datasets and in different geographical environments before anyone can assume the scores will hold elsewhere.
Angles, lighting and resolution
Field photos differ in viewing angle, lighting and image resolution. The researchers list all three as conditions still to be evaluated. A model trained on handheld inspection photos may behave differently on drone footage, zoomed-out overview shots or images taken under a deck in deep shade.
Not yet on mobile or edge hardware
Inspectors work on site, often with patchy connectivity. The study names mobile and edge-computing hardware as a future test. SegFormer-B5, the best Level 2 model, is the largest variant in its family, so running it quickly on a tablet or a drone’s onboard computer is a genuine engineering question rather than a formality.
Pattern is not capacity
A visual class does not tell an engineer how much steel is left. The team envisions linking the framework with inspection histories, structural information and physical measurements. Until that happens, the output describes the surface appearance of bridge corrosion, not the load-carrying capacity of the member.
What Bridge Owners Can Do With Bridge Corrosion AI Now
The framework is research, not a product, but the lessons transfer. Owners and inspection consultancies can prepare their data and processes so that tools like it are useful when they arrive.
Standardise how photos are taken
Consistent distance, angle and lighting make any model more reliable. Simple capture guidance, such as an overview shot plus a close-up of each defect with a scale reference, improves both human review and machine analysis of bridge corrosion. It also makes photos comparable across inspection cycles.
Keep inspectors in the loop
Use segmentation as a draft, not a verdict. An inspector who corrects a model’s map is producing new labelled data at the same time. That is how the 92.81% agreement standard in this study was reached: careful, repeatable human labels come first.
Pilot on archived reports first
Most owners already hold years of inspection photographs. Running a model over archived images, and comparing its bridge corrosion maps with what inspectors recorded, is a low-risk way to measure value before changing a live inspection process. It also shows where the model fails on your own structures. Teams exploring this can borrow ideas from predictive maintenance and analytics programmes.
Ask vendors for per-class scores
An average mIoU hides rare-class performance. Ask for IoU and F1 per corrosion pattern, a boundary metric, the countries and structure types in the training data, and results on an independent test set. The Saitama paper sets a good template for that list.
Treat inspection images as a governed asset
Inspection photographs of critical infrastructure deserve the same cybersecurity controls as other sensitive asset records. Clear ownership, access control and retention rules also make it easier to share data with research partners safely. Well-governed data is what lets an inspection programme train or validate its own models later, much like the manufacturing case in our look at an AI-powered inspection system for 3D printers.
Frequently Asked Questions About Bridge Corrosion AI
What is the two-level bridge corrosion framework?
It is a semantic segmentation approach from Saitama University. Level 1 marks every pixel of visible corrosion in a steel bridge photo. Level 2 assigns those pixels to one of four visual patterns: Uniform, Crevice, Underfilm or Other Localized corrosion.
Which models performed best?
SegFormer-B2 with boundary-aware supervision led Level 1, with 84.30% mIoU, 73.39% corrosion IoU and 84.65% F1. Mask2Former-Small had the best Boundary-F1 at 69.10%. SegFormer-B5 led Level 2 with 60.45% five-class mIoU and 67.81% mean F1.
How big were the datasets?
The binary dataset held 1,960 images and the categorised dataset 1,738 images, all taken from Japanese steel bridge inspection reports. Annotation agreement was 92.81%, with a Cohen’s kappa of 0.892.
Can it replace a bridge inspector?
No. It describes the visible pattern of bridge corrosion but does not measure section loss or judge structural safety. The researchers position it as more informative input for inspectors, and plan to link it with inspection histories and physical measurements.
Where was the research published?
The paper, “A two-level semantic segmentation framework for visual corrosion pattern assessment in steel bridges”, was published online on 5 August 2026 in Computer-Aided Civil and Infrastructure Engineering, by Sal Saad Al Deen Taher and Ji Dang.
References
A two-level semantic segmentation framework for visual corrosion pattern assessment in steel bridges
AI Framework Boosts Steel Bridge Corrosion Checks (Mirage News)
Structural Mechanics and Dynamics Group, Saitama University
MLIT Road Bureau: Road Maintenance
FHWA National Bridge Inventory Tables
SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers
Masked-attention Mask Transformer for Universal Image Segmentation
A multilevel bridge corrosion detection method by transformer-based segmentation in a stitched view
Quantifying element-level corrosion intensity in steel bridges using context-aware late fusion
Evaluation of AI-based Visual Crack Detection in Steel Bridges Using Probability of Detection
Computer Vision-Based Bridge Inspection and Monitoring: A Review
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.