Zero print defects without slowing the line

The problem

At line speed the eye tires, and sampling leaves gaps.

Those defective sheets move on to forming and end up as a defective can body or, worse, a varnish failure that compromises food contact. Sampling inspection leaves gaps.

• At real lithography line speed, the human eye tires and lets through sheets with color registration defects, scratches or varnish bubbles.
• Sampling inspection of 1-2% of sheets leaves every other sheet unchecked. A register shift that starts between two samples can run for hundreds of sheets before anyone sees it.
• A defective sheet moves on to forming and becomes a defective can body or, worse, a varnish failure that compromises food contact. Forming, welding and coding add cost to a sheet that was already lost.
• By the time the complaint arrives from the packer client, the run has been formed, coded and shipped. At that point the cost is no longer a sheet, it is a run and a client conversation. And the root cause is hard to find, because nobody knows which sheets of the run were affected.

How it fits the IRIS system

Edge — local CNN vision over the litho line, no image sent to the cloud.

Edge — local CNN vision, with no image sent to the cloud (latency, cost, bandwidth). Model trained on good and bad examples of the specific format and Pantone. Automatic rejection before forming.

The model is trained on good and bad sheets of your own formats and Pantones, not on a generic library. That is what lets it reject a register shift without rejecting the design. And it sees every sheet, not the one in fifty that a person has time to look at.

See the full IRIS architecture →

Before and after

Sampling the litho output versus inspecting every sheet

Aspect	Sampling inspection	With iLEAN Edge
Sheets inspected	1-2% sample	100%
Color registration defect	Caught if it falls in the sample	Rejected sheet by sheet, at line speed
Varnish bubbles and scratches	Depend on the inspector's fatigue	Classified at line speed
Where a bad sheet is stopped	At the packer client, as a complaint	Before forming
Image processing	—	Local, no cloud latency or bandwidth
Model reference	—	Your formats and Pantones

1-2% sampling → 100% sheet control. Print/varnish defect complaints → drastic reduction.

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 the current reject ratio.
• Inspection coverage goes from a 1-2% sample to every printed sheet. Every reject is stored with its image, which turns complaints into data per unit and per run. Quality can see whether a defect grows with a blanket, an ink batch or a specific unit of the press.
• Print and varnish defect complaints from packer clients drop sharply. Each one avoided is a client conversation that does not have to happen.
• And defective sheets are removed before forming adds the cost of the can body to them.

Estimated payback of 5 to 12 months depending on current reject ratio. Estimate to validate.

And the fair question from the production manager

“What if it rejects good sheets?” — the false reject is the real risk, which is why the model is trained on good and bad examples of the specific format and Pantone, under that line's lighting. Classifying a known defect type on a known design is an anchored task where the best models drop below 1.5% error [1], and borderline sheets are escalated to a person instead of being discarded. Every decision on a borderline sheet feeds back into training, so false rejects fall as the model learns your designs.

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