OCR of DOM/EXP on curved cosmetic packaging with AI — the blurry character reaches the shelf, and comes back as a claim.

The problem

The date is there, but the OCR cannot see it.

On a cosmetics packing line, the manufacturing date (DOM) and the expiry / period-after-opening date (EXP / PAO) are stamped on the packaging by inkjet or laser. The packaging is curved plastic, often pearlescent or metallic, and it moves at speed. Classic OCR, which learned to read flat, contrasted text, collides with three realities:

1. Reflection and curvature. Light bounces off the glossy plastic and deforms the character as seen from the camera. What on matte cardboard would be a trivial reading is, on a curved bottle, an angle lottery.
2. Low contrast. Inkjet on a pearlescent background leaves a diffuse stroke. Laser on a metallic background burns the plastic just enough to be visible — or not.
3. The SKU change. Every reference has its geometry, its stamp position and its date format. Manual calibration does not arrive in time and the first units after a changeover leave without effective verification.

The packing manager knows this, but when classic OCR starts firing "no read" on 5% of units, the plant ends up checking everything manually — and a human at 60 bottles/min does not scale. The classic system works 99% of the time. That 1% is what reaches the retailer as an illegible-date complaint or, worse, as an out-of-date unit on the shelf.

How it fits into the IRIS system

iLEAN Vision does not replace your inkjet — it seals the crack between what is stamped and what OCR can read.

The date-OCR problem is not a lack of information: it is information living in islands (the ERP's lot, the SKU's stamp position, the inkjet's configuration, the optical reading) that, at the critical moment (the bottle passing under the camera), is never cross-checked in time. iLEAN acts as the filler that closes those gaps without asking you to change the inkjet or the line.

Edge sees the date on the packaging. Connect reads the lot and the SKU's position from the ERP. The agent cross-checks with the print history and, if the reading does not fit, holds the bottle. The person signs off — never the other way round.

The iLEAN pieces applied to DOM/EXP OCR on curved packaging:

• iLEAN Vision (Edge) — a terminal with machine vision (CNN) trained on low-contrast stamps on curved plastic, with low-angle or coaxial lighting depending on the packaging. It reads, decodes the format and triggers the actuator (ejector, light signal) in milliseconds. It works with no network. If the plant loses its WiFi, Edge keeps reading and holding.
• Connect — captures from the ERP the lot, the expected manufacturing date and the PAO/EXP formula of each SKU. And it captures from the MES or the inkjet's PLC the actual print configuration (not what the sheet says, what is loaded in the machine).
• Agent — cross-checks the optical reading with the expected lot. If the average confidence drops (a degraded inkjet, a new supplier's plastic, a shifted position), it alerts before the first badly stamped unit reaches the retailer. The person validates and signs; the line never restarts on its own.

See the full IRIS architecture →

Before and after

Classic OCR vs. ERP-anchored OCR with iLEAN Vision

Aspect	Classic OCR + manual re-reading	With iLEAN Vision + Connect + Agent
Low-contrast stamping	Frequent "no read", discard by default	Plant-trained CNN reads even with poor contrast
Data validation	Blind optical reading	Cross-check with the ERP's lot, anchored OCR
SKU change	Manual zone and format calibration	The SKU's position and format read from the ERP
Degraded inkjet	Detected when the complaint arrives	Average confidence drop, early alert
Operation with no network	n/a	Edge keeps operating on cabinet power
EU 1223/2009 file	Reconstructed by hand per lot	Dossier per unit with a photo of the stamp, automatic

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 figures of your plant. We set it out so the committee has an order of magnitude; we refine it during the diagnostic.

• A cream line with pearlescent PE bottles, side date inkjet, multi-SKU with weekly format changes and a manual re-reading percentage that weighs on the OEE.
• iLEAN Vision pilot on one line (camera with lighting adapted to the packaging, actuator, integration with the ERP's SKU master). First expected value within a few weeks.
• Indicative payback between 4 and 9 months, depending on the current manual re-reading percentage and the frequency of retailer claims for illegible dates.
• A reduction of units with inconclusive OCR of ≥ 30% over the baseline — the hard lever is a single avoided retailer withdrawal. One serious debit note pays for the pilot.

And the quality manager's reasonable doubt

"What if the AI invents a date that was not stamped properly?" — hallucination is a problem of free generation, not of anchored tasks. When the AI simply recontextualises a data point anchored to the ERP's lot (the date this bottle should carry), the best models brought the error below 1.5% ^[1]. And even so, what is critical is never decided alone: iLEAN holds the bottle if confidence does not reach the threshold and the person signs off. The three safety rings are there for exactly this.

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