A coating defect is only visible while the piece goes past

The problem

A coating defect is visible for one second, and then it freezes.

A coating defect is superficial and visible, but only if somebody is looking at that piece at that exact instant. Sampling-based inspection leaves gaps by definition. And this product has a cruel particularity: once the piece enters the tunnel and freezes, it becomes indistinguishable from the rest. It is packed, palletized and distributed. The defect is discovered when the croquette opens up in the restaurant's fryer, and the complaint arrives with a photograph. Every defective piece that gets through also consumes tunnel freezing capacity and packaging, both of which are lost too.

• Incomplete breading, a piece burst in the pre-fryer, a misshapen one, one outside weight tolerance: all superficial, all visible — but only if somebody is looking at that piece at that exact instant. At the real speed of a croquette line, sampling leaves gaps by definition.
• This product has a cruel particularity: once the piece enters the tunnel and freezes, it becomes indistinguishable from the rest. It is packed, palletized and distributed.
• The defect is discovered when the croquette opens up in the restaurant's fryer, and the complaint arrives with a photograph and the distributor's name on it.
• Every defective piece that gets through also consumes tunnel freezing capacity and packaging, both of which are lost too.

How it fits the IRIS system

Edge — AI vision executed locally, piece by piece, before the tunnel infeed.

Edge — AI vision executed locally.

Rejecting before the tunnel instead of after the customer does two things at once: the scrap is never frozen and never packed, and every rejection is logged against the batch. That second effect — a defect history by reference and line — is the one that ends up changing how the breading machine is run, because for the first time the plant can see that bare pieces appear above a certain belt speed on one format and not on another. The supervisor's observation from case 4 gets a number behind it.

• An overhead camera on the breading outfeed belt, before the tunnel infeed.
• A neural network trained on good and bad pieces of the plant's real formats — croquette, large croquette, mini croquette, filled pepper, bite-size — rather than a generic model.
• Inference in milliseconds per piece on a local GPU: no image is sent to the cloud, so neither latency, bandwidth nor connectivity is a limit.
• The defective piece is rejected before the tunnel.
• Every rejection is logged against the batch and feeds the defect history by reference and line.

See the full IRIS architecture →

Before and after

Today's inspection versus Edge vision

Aspect	Sampling inspection	With iLEAN Edge
Coverage	A small percentage of pieces	100% of pieces
Where the defect is found	In the restaurant's fryer	On the breading outfeed
A burst or bare piece	Frozen, packed and shipped	Rejected before the tunnel
Tunnel capacity and packaging	Spent on scrap	Spent on good product only
Defect history	None	By reference, line and batch
Images sent to the cloud	—	None: inference on a local GPU

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 scrap ratio.
• The saving is made of scrap never frozen, packaging never used and complaints never received.
• From sampling a small percentage to inspecting every piece, at line speed, without slowing the breading line.
• And a defect history that shows, for the first time, which reference and which line speed produce the bare pieces.

estimated payback of 5 to 12 months depending on the current scrap ratio. The saving is made of scrap never frozen, packaging never used and complaints never received. *Estimate to be validated*.

And the fair question from the production manager

“What if it rejects good croquettes?” — the false positive is the real risk of any vision system, which is why the network is trained on good and bad pieces of your own formats — croquette, large croquette, mini, filled pepper, bite-size — under that belt's lighting, not on a generic model. Classifying a piece against a trained reference is an anchored task, where the best models drop below 1.5% error [1]. Borderline pieces are not simply rejected: they are escalated for a person to decide, and every decision feeds back into training.

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