OCR of batch and expiry on cosmetic aluminum tubes — curvature and glare cannot cost you a customer rejection.

The problem

Classic OCR was designed for flat labels — not for glossy aluminum tubes with low-contrast inkjet marking.

The cosmetics sector packs body cream, sun cream, hand cream and toothpaste in aluminum tubes for a sound reason: barrier, light weight, premium perception. But that same tube is the worst possible surface for classic OCR:

1. Curvature — the inkjet text is not flat; it distorts as soon as the camera moves off the optimal center.
2. Specular glare — glossy or satin aluminum reflects the illuminator straight back, and the camera sees more glare than character.
3. Low-contrast marking — light ink on a light background (white batch code on a pastel pink tube) or dark ink on a dark background (black batch code on a graphite tube) are the norm, not the exception.
4. SKU-to-SKU variability — every reference has its own tube, its own color, its own typeface, its own marking height. The classic OCR threshold that works for one SKU rejects the next one.

The result: false rejects (perfect tubes into the scrap bin, unnecessary line stops) and false positives (a tube with unreadable marking slips through to shipping and comes back as a B2B customer return). The operator ends up raising the threshold so the line keeps running, and the problem is masked until the complaint arrives.

How it fits the IRIS system

iLEAN Vision does not require changing the printer — it puts modern vision where classic OCR cannot reach.

The problem with OCR on aluminum is not the camera, it is the algorithm: classic OCR binarizes the image with thresholds and breaks down when contrast is low or glare is high. A CNN trained on hundreds or thousands of real tubes from your line learns the pattern even when the light is not perfect — and it improves every week with new samples. iLEAN Vision acts as the putty between your inkjet and your MES: nothing new on the line, reliable reading behind it.

Edge sees the tube in milliseconds. The agent cross-references the batch it read with the work order. If they do not match, it holds the tube. The person signs — never the other way round.

The iLEAN pieces applied to OCR of batch and expiry on cosmetic tubes:

• Edge Vision — a terminal with a CNN over the packing line, a camera with lighting designed for metal surfaces (diffuse coaxial, polarized if needed) and a millisecond reject actuator. It works with no network: if the plant loses WiFi, Edge keeps reading, validating and holding.
• Connect — captures the work order whether it comes from the MES, the ERP or a planning spreadsheet, and hands it to Edge at second zero. When the SKU changes, Edge already knows which batch code and which expiry date it should be seeing — and anything that does not match is held.
• Agent — cross-references the Edge reading with the work order and assembles the per-batch audit dossier: photo of the marking, batch read, batch expected, human decision on every reject. When the pharmacy retailer's questionnaire arrives, nothing has to be reconstructed by hand.

See the full IRIS architecture →

Before and after

Classic OCR on aluminum vs. cross-referenced reading with iLEAN Vision

Aspect	Classic OCR + threshold	With iLEAN Vision (trained CNN)
Glossy surface / glare	False rejects from excess light	Dedicated lighting + CNN robust to specular glare
Low-contrast marking	The threshold lets it through or rejects it at random	Learned from real samples of your line
SKU changeover	Manual parameter readjustment, line stop	Model change in seconds, no stop
Traceability of the reject	Counter on a screen, the image is lost	Photo of the tube + reading + human decision, stored
Operation with no network	n/a	Edge keeps running on cabinet power
Retraining after an ink change	Call the integrator, weeks	New samples, hours, without touching the line

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 line. We put it forward so the committee has an order of magnitude; we refine it during the diagnostic.

• Aluminum tube packing line for body cream, sun cream or toothpaste, multi-SKU, with frequent format changes and a current classic OCR that generates visible false rejects.
• Vision pilot on one line (Edge + camera with the right lighting + actuator + MES integration). First value expected within a few weeks: detection above 90% and a visible reduction in false rejects.
• Indicative payback between 4 and 9 months, dominated by B2B returns avoided + scrap recovered (good tubes that classic OCR was rejecting) + savings on line stops for threshold readjustment.
• The hard lever is a single large B2B customer return avoided: the cost of one return for unreadable marking usually pays for the entire pilot.

And the quality manager's reasonable doubt

“What if the CNN misreads a batch code and lets a bad tube through?” — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI merely recontextualizes a piece of data (reading the marking and comparing it with the work order), the best models brought error below 1.5% ^[1]. And even so, what is critical is never decided alone: in a doubtful case, Edge holds the tube and the person signs.

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