Fatigue cracks
At weld toes, radii and bolt holes, where the load concentrates. Segmented one by one, so two cracks on one bracket are two records.
Industries
Metal and infrastructure
Watch a crack grow instead of finding it after it has gone
Brackets, flanges and weld seams carry load until they do not. Today someone goes out with a UV lamp on an interval and looks. A fixed camera on the part reads the same coating every day, measures the indication, and tells you when it crosses your limit.
- Reads the Riluminati coating: segments and measures the indication
- Fixed camera, continuous watch, no routine site visit
- Alerts on a length or area limit you set
We support the inspection of mounting brackets, flanges, weld seams, load-bearing steelwork.
Case 01 · Monitoring infrastructure over time
Watching brackets in service, day after day
Not an inspection station — a camera left on a structure, reading the same bracket every day. The brackets are coated with Riluminati, so a crack in the steel shows yellow-green under UV. Riluminati makes it visible; we measure it and watch it grow. Deepvis does not produce or sell the Riluminati coating — it supports the inspection of it. It does not replace a certified NDT inspection. It watches the same indication in between them.
- Images to train
- 130
- Precision
- 100%
- mAP@50
- 90.9%
Photographs of cracked brackets, marked up by hand.
No false positives — everything it flagged was a real crack. Nobody gets sent out for nothing.
The standard benchmark score for how closely the outlines match the ones drawn by hand.
How crack luminescence works
- Top coat. Black, sprayed over the indicator layer.
- Indicator coat. Fluoresces yellow-green under 365 nm UV.
- Substrate. Steel or aluminium. A crack in it cracks the coating.
The coating system is Riluminati, developed by MR Chemie with BAM. Two cans: 815 indicator layer and 816 black overlay, both from MR Chemie. Deepvis supports the inspection of it.
Case 02 · Weld inspection on the line
Automated weld inspection at JP Group
This one is an inspection station in a production cell. Circumferential welds on machined steel, checked on the surface for pores and spatter as the parts come through. The point is not only finding the defect — it is who has to be standing there when it is found.
The bottleneck was a person, not a machine
- 01 An operator can run a set number of parts — say fifty — before the work has to be verified together with a welding coordinator.
- 02 Hit that number and production of that part stops until the coordinator is free. If they are off sick or tied up on another job, the cell waits.
- 03 Automating the weld control removes the gate. The operator keeps running, and the coordinator spends their time where a qualified judgement is actually needed.
ISO 5817
Working to the standard, not around it
A pore is not automatically a reject. ISO 5817 sets quality levels — B, C and D — and the limits scale with material thickness, so the same pore passes on one part and fails on another. Which level applies is on the drawing.
- That makes it a measuring job rather than a spotting job. We segment the imperfection and report a size, which is the number a level is held against.
- Pores and spatter are separate imperfection types with separate limits, so they are trained and reported as separate classes.
- Every part is imaged and kept, which is what makes the control auditable rather than a claim.
Read in blog
AI-løsning fjerner flaskehals på maskinfabrik – kan nu tage større ordrer
The case on video, in Danish.
“Deepvis fandt defekter, som operatøren og svejsekoordinatoren ikke kunne se med det blotte øje. NSK har set resultaterne og kan bekræfte dem.”
Coming soon · Ongoing MADE project
AI on the radiograph: X-ray weld inspection
Surface inspection scans the weld optically. X-ray reads what is inside the weld: porosity, inclusions and lack of fusion — none of it visible on the surface. We are training models to read radiographs in an ongoing demonstration project with MADE. Not a shipped capability yet — in active development and coming soon.
MADE Manufacturing Academy of Denmark, demonstration project Visit site → A weld on a metal part, seen through X-ray. The dark line along the root is the kind of indication only X-ray sees — nothing on the surface gives it away.
In partnership with
On the machine
Mounting, cycle time and rejects
Where the camera goes, how fast the result comes back, and what happens to a bad part.
- 01
Camera fixed on the part
Mounted on the bracket or the flange and left there. No access platform, no shutdown, and no trip out for a routine look.
- 02
Measured, not eyeballed
The indication is segmented, so you get a length and an area, and the same measurement method next month.
- 03
Alerts on your limit
Set the limit in mm or mm². You hear about it when the limit is crossed, not when someone next happens to visit.
- 04
Every reading kept
A growth curve per bracket, instead of a pass or fail per inspection round.
Defect list
What we inspect for
Crack growth
Length and area measured the same way every reading, so this week and last week are comparable numbers rather than two opinions.
Coating wear, not a crack
A top coat worn thin fluoresces as well, in diffuse patches. Telling that apart from a crack line is most of the job, and getting it wrong means a call-out for nothing.
What changes
What changes on the line
Four things a shift notices.
No routine site visit
Interval inspection costs access, a rope team or a platform, and usually downtime. A fixed camera reads the same part daily for the cost of the camera.
You can see the trend
A crack that has grown over a month is a different decision from one that has not moved in a year. An interval inspection cannot tell those two apart.
Replace on evidence
Brackets come out when the measurement says so, rather than on a fixed schedule or after a failure.
Feeds the next design
Where the cracks start and how fast they run, across every bracket you are watching. That goes back to the designer.
The method
How we find them
Four steps, whatever the part is. Everything above was marked up by hand first. That is how the model learns your defects, and the only way it learns them.
You set what counts
We start from your reject criteria, on your parts. The model returns what you marked up and nothing else.
Lighting first
Camera, optics and lighting are picked for the material. Frosted PP and a black closure are different jobs.
About 120 parts
A hundred frames off the running tool, 120 marked up. Your resin, your regrind, your colours, your fixture.
Two hours to a model
Training takes about two hours. After that every part is checked in the cycle and rejected on your limit.
First step
Prove the defects are detectable
You send 100 parts — 50 good, 50 faulty. We photograph them, train a model on them in our lab, and 14 days later you have the report.
What the report contains — click a page
Guarantee
If we cannot find the defect type you point at, you get your money back. Buy the system within 60 days and the fee is credited in full.
Book a detectability analysis
DKK 8,000
FAQ
Frequently Asked Questions
Simple and Easy to deploy AI Inspection Systems for modern production environments.
What types of defects can Deepvis detect?
Deepvis can detect complex and unpredictable defects that are common in recycled plastic, such as black specks, short shots, overmoulding, contamination, burn marks, and geometry defects. The model is trained on annotated examples of the defect types you care about, and holds up against the material, colour and surface variation that comes with recycled feedstock.
Do I need thousands of images to train the system?
No. Around 120 annotated images is enough to train a model. This is a large reduction compared to traditional systems that require thousands of manually labelled samples. It is however important to note that challenging applications typically perform better, the better the data.
Can Deepvis keep up with fast production speeds?
Yes — with the right GPU hardware, Deepvis can analyze parts in real time down to 0.5-second cycle times, making it suitable for high-speed packaging and manufacturinglines.
How is Deepvis different from traditional vision systems?
Traditional systems rely on rule-based image processing and require reprogramming every time a product changes. Deepvis uses AI to adapt to product variation and requires no specialist for setup, making it ideal for agile manufacturers where long and complex setuptimes are out of the question.
What's included in the Deepvis inspectionsystem?
Each unit includes a camera, AI edge computer, and software platform for inspection. Customers can purchase the full solution as a one-time investment, with optional software licensing and support packages.
How do I know if Deepvis is right for my application?
If your quality control process is manual, time-consuming, and affected by frequent changes in product, lighting, or material, Deepvis is built for you. Our system is designed to reduce man-hours in visual inspection — especially in high-volume, high-variation environments where traditional systems fall short. We typically start with a discovery call and a pilot to demonstrate time savings and operational impact.
Explore Smarter Quality Control with Deepvis
Discover how AI-driven visual inspection can reduce errors, cut costs, and adapt to your production line — without complex setups.