The Paint Defect That Never Reaches the Oven

The problem

A defect found after the oven costs a repaint; before it, far less.

100% human visual inspection at line cadence is impossible to sustain shift after shift; sampling-based inspection leaves gaps. A defect caught after the curing oven costs far more to fix (full repaint) or reaches the customer directly as a perceived-quality claim.

• Inspecting 100% of every body's surface at paint-line cadence is not humanly sustainable shift after shift. The eye tires; the line does not slow down.
• Inspection usually happens after the curing oven, by sampling or in the light tunnel, where a dirt inclusion, a run or a crater is already baked into the clearcoat.
• Caught there, the unit needs sanding and spot repair, or a full repaint loop through the booth and the oven again.
• Missed there, it reaches final assembly or the customer as a perceived-quality claim on a vehicle that is otherwise finished.

How it fits the IRIS system

Edge at the booth exit — infer in milliseconds, divert before curing.

Edge: an industrial camera at the cabin exit plus a local GPU run a CNN trained on this plant's actual good/bad finish examples, inferring in milliseconds and diverting the defective unit before the oven.

Moving the eye ahead of the oven changes the economics of the defect: the same inclusion that costs a repaint after curing costs a local repair before it. And because the inspection is local and runs on every body, the plant finally has defect data by booth, color and shift instead of a sample. Paint engineering stops arguing about where the dirt comes from and starts looking at the map.

See the full IRIS architecture →

Before and after

Post-oven sampling versus Edge at the booth exit

Aspect	Today	With iLEAN Edge
Point of inspection	After the curing oven	At the paint booth exit
Body coverage	A sample, or what the light tunnel catches	100% of the bodies
Repair for a caught defect	Sanding, spot repair or full repaint	Local repair before curing
Defect data by booth and color	Partial	Every body classified
Where the paint inference runs	Nowhere: there is no inference	Local GPU next to the booth
When the unit is diverted	After it already cost an oven pass	Before the oven

Partial post-oven sampling → 100% body coverage before curing.

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 your current repaint and rework ratio.
• 100% body coverage before curing, against today's partial post-oven sampling, on every shift and at full line cadence.
• Fewer full repaint loops, each of which consumes booth time and oven capacity the line needs for new units.
• And defect classification by booth, color and shift that turns rework into a cause you can actually attack.

Estimated payback 5-12 months depending on current rework ratio. *Figure to be validated*. (Payback 5-12 months · 100% control before curing)

And the fair question from the production manager

“What about false rejects that send good bodies off the line?” — classifying a known defect catalog on your own colors under fixed booth-exit lighting is an anchored vision task, where the best models drop below 1.5% error [1]. The model is trained on this plant's real good and bad bodies, and borderline cases are not diverted blindly: they go to the inspector at the console, who decides, and that decision feeds the training so the same borderline case is handled better next time.

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