On prepainted steel the defect cannot be repaired: it is scrapped. And today your customer finds it

The problem

At line speed the human eye does not inspect: it samples.

A scratch, a roll mark, an embedded particle or an oil stain turns a prepainted strip into second-grade material. The surface is the product and it admits no repair. The same goes for edge burr on the slitter, which cannot be judged by eye at line speed and which the customer discovers when it breaks their die. And the worst part is where it gets discovered: the defect travels coiled inside the output coil and is not seen until the customer unwinds it in their press. By then the cost is no longer the material: it is sorting on somebody else's premises, return freight and trust.

• A longitudinal scratch, a roll mark, an embedded particle or an oil stain turns a prepainted strip into second-grade material. The surface is the product and it admits no repair.
• Edge burr on the slitter cannot be judged by eye at line speed, and the customer discovers it when it breaks their die.
• One or two percent inspected means ninety-eight percent leaving unseen, coiled inside the output coil where nobody can look.
• The defect is found when the customer unwinds it in their press. By then the cost is no longer the material: it is sorting on somebody else's premises, return freight and trust, and the whole coil is under suspicion rather than the ten meters that were actually bad.

How it fits the IRIS system

Edge — local inference on a GPU inside the plant, meter by meter, before the strip is recoiled.

Edge — CNN vision with local inference on a GPU inside the plant, without sending images to the cloud: latency, cost and bandwidth solved at the root. Models are trained on real samples of the material this plant cuts: longitudinal scratch, roll mark, embedded particle, stain, rust, missing coating and edge burr. The system records the exact meter position of the defect and stops the line when the defect is continuous. The final decision — trim, mark or downgrade — stays with the operator.

The models are trained on samples of the material this plant cuts, not on a generic catalog, and the images never leave the plant: latency, cost and bandwidth solved at the root. The system records the exact meter of every defect and stops the line when the defect is continuous; the decision to trim, mark or downgrade stays with the operator, who now decides on evidence instead of on a glance.

See the full IRIS architecture →

Before and after

Today's inspection versus Edge vision

Aspect	Visual sampling	With iLEAN Edge
Coverage	1-2% of the strip	100% of every linear meter
Where the defect is found	In the customer's press	At the slitter or leveler exit
Position of the defect	Unknown: somewhere in the coil	The exact meter, before recoiling
Edge burr	Judged by eye, if at all	Classified at line speed
A continuous defect	A whole coil downgraded	The line stops
Images	—	Never leave the plant

from 1-2% sampling to 100% control of every linear meter produced. From defects found by the customer to defects found on the line, with the exact meter located.

Impact estimate

Impact estimate — to be validated with your numbers.

The block below is an estimate to be validated against your plant's actual data. We put it forward so the committee has an order of magnitude; we refine it during the assessment.

• Estimated payback 5-12 months, depending on your current downgrade rate and the weight of prepainted in the mix.
• A significant reduction in surface claims, because the defect is found on your line instead of in the customer's press.
• From 1-2% sampling to 100% control of every linear meter produced.
• And sorting, return freight and the conversation about trust stop being part of the cost of a scratch: the bad meters are trimmed or marked in your bay and the rest of the coil ships as first grade.

estimated payback 5-12 months depending on your current downgrade rate and the weight of prepainted in the mix, with a significant reduction in surface claims. *An estimate to be validated*.

And the fair question from the production manager

“What if it downgrades good material?” — the false positive is the real risk of any vision system, which is why the model is trained on real samples of your own prepainted and galvanized material under that line's lighting, not on a generic model. Telling a scratch from an oil stain from a reflection on a known surface at a fixed camera position is an anchored task, where the best models drop below 1.5% error [1]. And a borderline detection does not downgrade anything: it marks the meter and the operator decides on the evidence.

^[1] OpenAI paper "Why Language Models Hallucinate", 2025 — on the reliability of AI in anchored tasks.