Glass and metal detection in packing with AI — one foreign body on the shelf is a mass recall.

The problem

The monolithic sensor sees a lot, but it does not see everything. And a single foreign body is a recall.

The metal detector has been installed on every serious line for decades. It works well with steel, worse with thin aluminum, poorly with materials that are not magnetic. The X-ray widens the range — it captures densities — but loses contrast when the contaminant is the same material as the container (glass on glass, a small bone on soft product). And then there are the cases no monolithic sensor covers:

1. Plastic fragments from the supplier's own packaging, especially if they are black or opaque — the X-ray does not detect them well because of their density, but a vision system does see them.
2. Organic foreign bodies (cardboard, thread, PPE fragments) that slip in from the manual loading area and reach packing without anyone noticing.
3. Color or texture anomalies that are not a contaminant but are a warning sign — a darkened area that could be the start of mold, a small stain that could be rust from the equipment.
4. The calibration test piece that nobody ran through this morning because the operator was in a hurry — and the detector spent two shifts with its sensitivity down without anyone knowing.

The quality manager knows this. But they cannot be on every line, on every shift, looking at every tray. The classic system works 99.9% of the time. That 0.1% is the shelf recall, the front page of the local newspaper and the retailer's call canceling the contract.

How it fits the IRIS system

iLEAN does not replace your X-ray — it plugs it into the metal detector, a vision camera and a brain that cross-checks all three signals.

The problem is not a lack of sensors; it is that the sensors do not talk to each other. The metal detector puts out one signal. The X-ray another. A camera could put out a third, but it lives in another world. The death sentence of monolithic systems is that each one covers its own range and leaves the gap to the one next to it. iLEAN acts as the putty that seals those gaps without asking you to change your X-ray or your detector — it integrates them and adds the missing piece.

Edge watches the line with machine vision. Connect reads the signal from your X-ray and your metal detector. The agent cross-checks the three signals against the non-conformance history and fires the ejector before sealing. The person signs off — the line does not restart on its own if there is any doubt.

The iLEAN pieces applied to glass/metal contaminants in packing:

• iLEAN Vision (Edge) — a machine-vision terminal (CNN) over the packing belt, trained on samples of your product in your container. It detects what the X-ray misses for lack of density contrast: plastics the color of the product, organic fragments, surface anomalies. It fires the actuator (pneumatic ejector, signal light) in milliseconds. It works without a network.
• Edge legacy + Connect — Connect reads the signal from the X-ray and the metal detector whether it comes from a modern PLC, from the old local PC in the cabinet, or from an analog panel that only emits a dry contact. The capture level adapts to each machine; it does not force you to change the line.
• Agent — cross-checks the vision, X-ray and metal-detector signals against the non-conformance history of the batch, the supplier and the SKU. It tells a false positive (irregular container, wrinkled label) from a real case, prioritizes the doubtful cases for human review and keeps the file live for IFS/BRC without the quality team rebuilding anything by hand. The three safety rings guarantee that automatic rejection is reviewable and that what is critical is still signed by a person.

See the full IRIS architecture →

Before and after

Isolated sensors vs. cross-checked detection with iLEAN

Aspect	Stand-alone X-ray + metal detector	With iLEAN Vision + Edge + Agent
Contaminant coverage	Metal and high densities	Metal + glass + opaque plastic + organic
Glass in glass containers	Limited sensitivity	Dual-energy X-ray + vision trained on your product
False positives	Each sensor generates its own, they add up	The agent cross-checks the three signals and filters out the noise
Shift calibration	Paper checklist — sometimes signed without being done	Test-piece run logged with photo, alert if it fails
In-line rejection	Yes, but without detailed traceability of why	Ejector + photo + batch + reason, all in the file
IFS/BRC file	Rebuilt after the audit	Always ready — the auditor opens the screen and sees it

Impact estimate

Impact estimate for your plant — to be validated with your numbers.

The block below is an estimate to be validated with the specific data of your plant. We put it forward so the committee has an order of magnitude; we refine it during the diagnostic.

• Multi-SKU product packing line (sauces, preserves, dairy), 10,000-50,000 units per shift, with X-ray and metal detector already installed.
• Edge pilot with a vision camera over the belt + Connect reading signals from the existing sensors + agent. First value expected within a few weeks.
• Expected reduction of undetected foreign bodies ≥30% over the baseline (on many lines much more, but that is the defensible floor). Reduction of false positives through signal cross-checking: operators freed up for tasks that require judgment.
• Indicative payback between 4 and 9 months. The hard lever is a single recall avoided: product retrieved from the shelf, reverse transport, destruction, brand damage, retailer penalty. One recall pays for the pilot.

The quality standard the end customer expects

The automotive sector has been operating for decades to a demanding standard on the order of 25 PPM (parts per million) ^[1]. Food is not automotive, but the direction is the same: bringing defects down to figures where the monolithic sensor no longer reaches and only a cross-checked system can operate. The most demanding sector sets the standard; the rest, sooner or later, follow it.

And the quality manager's reasonable doubt — "what if the AI gets it wrong and lets a contaminant through?" — has an answer: hallucination is a problem of free generation, not of anchored tasks. Recognizing a glass fragment in an image is exactly an anchored task. In this type of task, the best models brought error below 1.5% ^[2]. And even so, automatic rejection stays traced and reviewable; the doubtful cases are signed by the person.

^[1] Symestic — quality standard on the order of 25 PPM in automotive.

^[2] OpenAI paper "Why Language Models Hallucinate", 2025 — on the reliability of AI in anchored tasks.