An image model can look at a photograph of a ballast tank and tell you, with reasonable reliability, that there is coating breakdown in the upper third of the frame. What it cannot tell you is whether that plate has lost twenty per cent of its as-built thickness — because that is a physical measurement, not something visible in a photograph, and it is the number on which the class decision actually turns. Holding both of those facts at once is the whole of a sensible position on artificial intelligence in vessel inspection. The detection capability is real and improving quickly: research systems now report crack detection accuracy around eighty-five per cent under genuinely difficult conditions including existing corrosion, weld seams and marine fouling, and the major classification societies have run their own image-recognition programmes to identify and measure coating breakdown from hull photographs. Equally real is the boundary. Surveyors have not been replaced, close-up visual inspection within arm's reach remains mandated for the vessels that matter most, and thickness gauging remains a contact measurement. This page sets out what image models detect, how they quantify it, where the hard limit sits, what the surveyor still decides, and how this capability fits into inspection software rather than standing apart from it.
TECHNOLOGY GUIDE · AI IN VESSEL INSPECTION
AI Defect Detection and Corrosion Analysis in Ship Surveys
What image models genuinely flag from inspection photographs — corrosion, coating breakdown, cracks — how they quantify extent, and the precise point at which a human surveyor still has to make the call.
HumanCapture
ModelDetect and measure
HumanReview and confirm
HumanDecide and action
What Image Models Actually Detect
The defect categories computer vision handles well are those with consistent visual signatures. This is a genuinely useful list rather than a speculative one, drawn from working systems rather than roadmaps.
Coating breakdown, including the early stages before visible rust forms
Corrosion patterns, with extent mapped across a surface rather than merely noted
Cracks, detected even against corroded backgrounds, weld seams and fouling
Marine fouling and growth affecting hull performance
Dents, deformation and physical damage to structure
Missing or wasted anodes
Propeller damage and blocked sea chest openings
Gaps in inspection coverage where an area was never adequately imaged
That final category is quietly one of the most valuable and least discussed. A model that knows which areas it has not seen turns an inspection into a measurable exercise — you learn not only what was found but what was never looked at, which is a question paper-based surveys have never answered well. To see which of these categories apply to your own vessel types and inspection photography, arrange a technical walkthrough or upload sample images in a trial.
Three Levels of Analysis
Not all image analysis is the same, and the distinction matters commercially because only the third level produces numbers you can put into a repair scope.
Level 1
Classification
The model states whether an image contains a defect type at all. Useful for sorting thousands of frames into those worth a human look and those that are not, but it says nothing about where or how much.
Level 2
Detection
The model locates the defect within the frame and draws a box around it, assigning a category and a confidence score. This is where reported accuracy figures — such as the roughly eighty-five per cent achieved by recent hull crack systems — generally apply.
Level 3
Segmentation
The model outlines the affected region pixel by pixel, producing a corrosion map. Mask area measured against total surface area yields the percentage of a surface affected, and where the image scale is known, that converts into real-world units.
Segmentation is what turns a photograph into data. A surveyor's note reading moderate corrosion in way of the upper stringer is a judgement that varies between individuals; a segmented map stating that eleven per cent of an examined surface shows coating breakdown is a figure that can be compared against the same surface twelve months later. Consistency over time is arguably the stronger benefit — not that the model is more accurate than an experienced surveyor on any single assessment, but that it applies the same standard every time, which is precisely what trend analysis requires and what human assessment struggles to deliver across different inspectors and years.
Put a real ballast tank photo in front of it
The fastest way to judge whether this is useful for your fleet is to run your own inspection images through it and see what gets flagged, what gets missed, and how the confidence scoring behaves on the marginal cases. That session takes about half an hour.
The Hard Limit: Photographs Do Not Measure Thickness
Any honest treatment of this technology has to be explicit about what images cannot do, because the gap between surface appearance and structural condition is where over-reliance becomes dangerous.
Class decisions rest on numbers a camera cannot produce
Class rules generally require immediate steel replacement when a plate shows a twenty per cent or greater reduction from as-built thickness, with substantial corrosion flagged as wastage approaches the allowable limit. Those are thickness values, obtained by gauging. A photograph shows the condition of a surface — it does not reveal how much material remains beneath it, and heavy visible corrosion on a thick plate may matter less than modest corrosion on one already close to its limit. Image analysis therefore informs where measurement should be concentrated; it does not substitute for the measurement itself.
Several related boundaries follow from the same principle. Coatings conceal what lies underneath, which is why techniques such as alternating current field measurement exist specifically to find cracks through coatings, and why acoustic methods are used for quantitative thickness readings through coatings. Enhanced survey programme requirements mandate close-up visual inspection with the surveyor within arm's reach for selected structures on bulk carriers and oil tankers — a requirement written around human proximity and tactile assessment, not photographic coverage. And image quality governs everything: poor lighting in a tank, water on the lens, obscuring fouling or an oblique angle degrade model performance in ways that are not always obvious from the output, which is one reason confidence scores need to be read rather than ignored.
Where the Surveyor Still Decides
The productive framing is division of labour rather than replacement. The model is an assistant that highlights exceptions; it is not a final authority on survey judgement.
The model does this well
The surveyor must do this
Processes thousands of frames consistently without fatigue
Sets the inspection mission and decides what needs examining
Flags candidate defects with location and confidence
Confirms or dismisses critical findings on the evidence
Quantifies affected area consistently across inspections
Judges whether a finding is class-relevant and what rule applies
Compares this inspection against previous ones on the same surface
Interprets structural significance in context of the vessel's history
Identifies areas not adequately covered
Approves actions and manages the repair decision and scope
Read the right-hand column carefully and a pattern emerges: everything there requires either authority or context that no image contains. Whether a finding is class-relevant depends on rules and on the vessel's survey history. Whether it warrants immediate repair depends on thickness readings, trading pattern, time to next drydock and commercial circumstance. The model narrows the field and brings evidence forward earlier; the decisions remain where they were. To discuss how that division would work across your own survey process, book time with our team or start evaluating it yourself.
What a Useful Finding Looks Like
A detection that arrives as an unlabelled box on an image is of limited value. What makes AI output usable in a repair scope, a class discussion, a claims file or a trend comparison is the metadata attached to it.
The trend dimension is where the commercial return concentrates. A single inspection tells you the current state; a consistently measured series tells you the rate of deterioration, which is what supports genuine planning — arriving at a drydock with a known repair scope rather than discovering it once the vessel is already in the yard, when scope changes are at their most expensive. It also underpins the shift toward condition-based rather than calendar-based maintenance that classification societies have openly described as the direction of travel.
Where This Sits in Inspection Software
Image analysis delivers very little as a standalone tool. Its value comes from being embedded in the system that already holds the inspection, the component register and the corrective action workflow.
Attached to the inspection that produced it
Photographs captured during an inspection are analysed within that record, so a detection is already bound to the vessel, date, inspector and checklist item rather than sitting in a separate image library someone has to reconcile.
Feeding the defect and corrective action workflow
A confirmed detection becomes a tracked defect with an owner, a deadline and evidenced closure — which is the step that converts an interesting observation into a managed item, and the step most standalone analysis tools simply do not have.
Building component-level history
Findings accumulate against the specific component rather than the inspection, so the next surveyor examining that structure sees what was found before and how it has developed.
Surfacing fleet-wide patterns
Consistent classification across vessels makes it possible to see that a coating system is underperforming across sister ships, rather than treating each finding as a local problem.
Supporting the human review step
Detections are presented for confirmation or dismissal by a qualified person, with that decision recorded and attributed — so the audit trail shows a human accepted the finding, which matters when the record is later examined.
Working offline where inspections happen
Images are captured in tanks and holds with no connectivity, so capture must work offline with analysis applied when the record syncs, rather than requiring a live connection at the point of inspection.
That last point deserves emphasis because it is frequently overlooked in demonstrations conducted in an office. The photographs this technology depends on are taken in exactly the spaces where no signal reaches, so the capture side has to function entirely offline regardless of where the analysis itself runs. To see how capture, analysis, review and corrective action connect in one workflow, request a workflow demonstration or trial the full loop on your own data.
Frequently Asked Questions
How accurate is AI corrosion and crack detection?
Accuracy depends heavily on defect type, image quality and conditions. Recent research systems for hull crack detection report accuracy in the region of eighty-five per cent under deliberately difficult conditions including existing corrosion, weld seams and marine fouling, which is a meaningful result given how hard those backgrounds are. Coating breakdown and corrosion detection are generally easier problems than cracks because the visual signature is larger and more distinctive. What matters practically is less the headline figure than how the system behaves on marginal cases — whether it reports confidence honestly, whether it fails toward flagging for review rather than silently dismissing, and how it performs on your own photography rather than on a research dataset captured under favourable conditions.
Can AI replace a surveyor?
No, and the sensible framing is that the model is an inspection assistant highlighting exceptions rather than a final authority replacing survey judgement. The division is reasonably clear. The model processes large volumes of imagery consistently, flags candidate defects with location and confidence, quantifies affected area, compares against previous inspections and identifies coverage gaps. The surveyor sets the inspection mission, confirms or dismisses critical findings, judges whether a finding is class-relevant and which rule applies, interprets structural significance in the context of the vessel's history, and approves the repair decision and scope. Everything in that second list requires authority or context that no image contains. Close-up visual inspection with the surveyor within arm's reach also remains mandated for selected structures under enhanced survey programme requirements.
Why can't image analysis assess structural condition on its own?
Because class decisions rest on thickness, and thickness is not visible in a photograph. Rules generally require immediate steel replacement when a plate shows a twenty per cent or greater reduction from as-built thickness, with substantial corrosion flagged as wastage approaches the allowable limit — values obtained by gauging, not by looking. A photograph shows the condition of a surface but not how much material remains beneath it, so heavy visible corrosion on a thick plate may be less significant than modest corrosion on one already near its limit. Coatings compound this by concealing what lies underneath, which is precisely why techniques exist to detect cracks and measure thickness through coatings. Image analysis is therefore best understood as directing where measurement should be concentrated.
What is segmentation and why does it matter?
Segmentation is the most useful level of image analysis because it outlines an affected region pixel by pixel rather than simply drawing a box around it, producing what amounts to a corrosion map. Measuring the mask area against the total surface area yields the percentage of a surface affected, and where the image scale is known that can be converted into real-world units. This matters because it turns a subjective description into a comparable figure. A note reading moderate corrosion varies between surveyors and years, whereas a stated percentage of affected surface can be compared directly against the same structure at the next inspection. The consistency is arguably more valuable than the accuracy on any single assessment, because trend analysis depends on the same standard being applied every time.
What makes an AI finding actually useful in practice?
The metadata attached to it. An unlabelled box on an image has limited value; a finding becomes usable in a repair scope, a class discussion, a claims file or a trend comparison when it carries the hull zone and vessel side, the frame area tying it to the ship's drawings, the specific component, the defect category, a confidence level, and the timestamp and inspection pass. Confidence in particular should be read rather than ignored, since it identifies which detections warrant human review rather than acceptance. Beyond the individual finding, the system should attach detections to the component's own history so accumulated findings show development over time, and should feed confirmed detections into the defect and corrective action workflow so they become managed items rather than observations.
Where does this technology add commercial value?
Principally in seeing condition earlier and arriving at a drydock with a known repair scope. Yard time becomes expensive when findings appear late, because a coating issue, propeller damage, blocked sea chest, rudder problem, worn anode or structural concern discovered after the vessel is already docked changes the scope at the worst possible moment. Consistent measurement across inspections also supports the shift toward condition-based rather than calendar-based maintenance that classification societies have described as the direction of travel, and gives a defensible basis for planning coating renewal rather than guessing. The secondary benefit is coverage assurance — knowing which areas were not adequately imaged answers a question that paper surveys have never handled well and that matters when a finding later emerges in an area nobody examined.
Judge It on Your Own Photographs
Vendor demonstrations use favourable imagery. The only assessment worth acting on is what happens when your own ballast tank and hull photographs go through the model — what it flags, what it misses, how confidence behaves on the marginal cases, and whether the findings arrive attached to the component, the inspection and the corrective action workflow rather than in a separate tool. That is a half-hour conversation, and it will tell you more than any accuracy figure.
