Visual control that never tires at industrial speed

The problem

At bagging speed, inspection is done by sampling.

At industrial speed, 100% visual control is done by sampling, letting occasional sealing, palletizing or label defects reach the industrial bakery customer.

• A bagging line fills and seals bags faster than any person can check them one by one, so visual inspection is done by sampling. One bag in fifty, or a check every half hour, is the norm.
• Between samples, the occasional defect passes: an open valve or bad seal, a crooked or missing label, a deformed bag that will break the pallet. None of them is frequent; all of them are expensive when they reach the customer.
• Those defects reach the industrial customer, who sees flour spilled on the pallet or a lot code they cannot read.
• Every one of them is a complaint, a return or a credit note, and a mark against the mill in the customer's supplier rating. That rating weighs on the next annual contract.

How it fits the IRIS system

Edge — a CNN trained on your bags, inference in milliseconds.

Edge camera over the line + a CNN trained on good/bad examples of that specific bag. Inference in milliseconds, automatic ejection of the defective unit.

The model is not generic: it is trained on good and bad examples of the specific bag on that line, with that lighting. That is what keeps false rejects low enough for the line to accept it. A system that rejects good bags gets switched off within a week; this one is tuned so that does not happen.

See the full IRIS architecture →

Before and after

Bag inspection today versus Edge vision

Aspect	Today	With iLEAN Edge
Coverage at the flour mill	Sampling	100% of bags
Valve or seal failure	Found on the pallet	Ejected before palletizing
Missing or unreadable lot code	Found by the customer	Rejected in line
Deformed bag	Breaks the pallet pattern	Diverted
Inspector fatigue at the flour mill	Grows over the shift	None
Evidence per reject	None	Image stored with the batch

From sampling-based control, to 100% unit control at line speed, with no fatigue.

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 5-10 months, depending on the volume of the bagging line.
• Fewer returns and complaints from industrial customers caused by sealing, palletizing or label defects. The defect is removed before the pallet is strapped, not after it is delivered.
• From sampling-based control to unit control at line speed, with no fatigue. The camera at the end of the night shift sees exactly what it saw at the start.
• And an image of every rejected bag stored with its batch, useful when a complaint has to be answered.

Estimated payback of 5-10 months depending on line volume, from fewer returns and complaints from industrial customers. Estimate to validate.

And the fair question from the production manager

"What if it rejects good bags and slows the line?" — the false positive is the real risk of any vision system, which is why the model is trained on your own bags. Classifying a seal or a label against known examples is an anchored task, where the best models drop below 1.5% error [1]. Borderline cases go to a separate lane for a person to decide, and each decision feeds the training. The reject rate is tracked from day one so the team can see whether the camera earns its place.

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