Services / 02

Inspection & Vision AI

Computer vision that finds cracks, corrosion, spalling and defects in photographs and video, so a qualified inspector's time goes to the places that need it.

Who this is for

The problem, stated plainly

Inspection is expensive because of the search, not the judgment. Nobody knows in advance which thirty percent of a facade has deteriorated, so a hundred percent of it gets scaffolded and tested. Nobody knows which welds on a line will fail, so every part gets the same look from an inspector whose attention is not the same at 4 p.m. as it was at 9 a.m.

Vision AI does not replace the judgment. It narrows the search: a first pass over every photograph, every frame, every part, that says where to send the person whose judgment matters. That reframing is what makes the economics work, and it is also what keeps the results defensible.

What we build

Facade and building inspection from photographs

Smartphone or drone photographs in, a ranked defect list out: cracks, tile lift, spalling, rust staining, water leak traces, each with a severity band and an urgency score so the report tells you what to do this quarter, not just what exists. The production version is the hosted facade inspection service on Airffic World; we also tune the same models to a client's own building stock and reporting format.

Infrastructure inspection from drone and camera footage

Bridges, transmission towers, tunnels and plant: detection of cracking, corrosion and deformation from drone imagery and fixed cameras, with the veteran inspector's criteria encoded into the model so the output is consistent between inspectors, seasons and years. Change detection against the previous survey is usually the most valuable feature.

Visual quality inspection for manufacturing

Line cameras and a model trained on your own defect history: scratches, foreign matter, color variation and shape defects flagged in real time. We start from the defect images you already have, which is why accuracy comes quickly, and we design the human review loop so the model keeps learning from what the line rejects.

Automated inspection reports

Every detected defect classified by location, type and severity and written straight into your report template, with the source image and the model's confidence attached. Report writing is where inspection hours disappear; this is where most clients see the first return.

Where the aviation model shows

Aircraft maintenance does not strip the airframe; it follows the system that showed an indication. That is the inspection philosophy here: a first pass that indicates, a qualified person who confirms. It also means we are deliberate about the limits, because in aviation a tool that overstates its certainty is more dangerous than no tool.

The reasoning is set out in facade inspection without scaffolding, including where AI screening does and does not belong.

How an engagement runs

StageWhat happensWhat you get
Free consultation (30 min)You describe the assets and the current inspection routineA view on whether photographs or video can carry the first pass
Sample assessmentYou send 20 to 50 representative imagesA first read on detectability, at no cost
Workflow improvement, implemented (from ¥50,000, about 2 weeks)Model tuned on your images and defect definitionsMeasured precision and recall on your data, and a sample report
Whole-operation architecture, implemented (from ¥300,000)Integration with cameras, drones or field apps; report automationA working inspection pipeline your team runs
Operate and improve (from ¥20,000 / month)New defect classes, seasonal recalibration, model monitoringAccuracy that holds as the asset base changes

Figures are indicative. If your case is a standard facade survey, the hosted service in Airffic World may be all you need, and we will say so.

What this will not do

A photograph shows the surface. Subsurface delamination, hidden corrosion and structural capacity are not in the image, and no model recovers information that is not there. Vision AI is a first-pass screen and a consistency tool; it does not replace a statutory inspection or a detailed survey by a qualified engineer, and every report we produce says so on its first page.

Frequently asked questions

How many images do you need to start?

For a first read, 20 to 50 photographs that include the defect types you care about. For a proof of concept, a few hundred labeled examples per defect class is comfortable; fewer is workable if the defects are visually distinct. We help with labeling and are honest when the data is not enough yet.

Will the model work on our specific buildings or products?

General models detect common defects well; tuning to your material palette, lighting and reporting categories is what a PoC is for. We measure precision and recall on your images, not on a benchmark, and we report both numbers rather than a single accuracy figure.

Can this be used for statutory or regulatory inspections?

As a screening and documentation aid, yes. As a replacement for the qualified inspector the regulation requires, no. The workable pattern is that the AI decides where the qualified person looks first, and the report records both the screen and the confirmation.

What about drone footage rather than still photographs?

Video is frames; the models work the same way and change detection across surveys is often easier because the flight path can be repeated. For drone-based inspection the operational side matters as much as the model, and we design both.

Send sample images or book a consultation

Twenty photographs is enough for a first opinion, free.

← All services