Every painted part inspected at the curing oven exit, with local CNN vision
At real line speed the human eye cannot inspect 100% of the parts on every rack: dirt inclusions, craters or sags get through and are discovered at assembly — or at the customer. iLEAN Edge infers in milliseconds per part at the oven exit, pulls the defective one before assembly and classifies every defect by type, position and paint lot.
FTQ defines the cost of the area — and it is managed by looking at one part in a hundred.
In a high-volume paintshop, FTQ is the metric that rules: it defines scrap, it defines rework and it ends up defining the conversation with the customer. And today it is managed with sampling-based human inspection, which has two limits that are nobody's fault:
- The eye tires and the line speed does not drop — inspecting 100% of the parts on every rack at real cadence is not humanly possible. A fraction gets looked at, and the fraction that does not still leaves for assembly.
- The defect is discovered where it is already expensive — a dirt inclusion or a crater that passes the filter shows up at assembly, when the part already carries added value, or at the customer, when it also drags a complaint behind it.
- Sampling generates no data to attack the cause — you know how much is rejected, but not in a structured way of what type, in which rack position or with which basecoat lot. Without those axes, improvement runs on intuition.
FTQ ends up being managed with the result in hand and without the variables that explain it, which is like trying to correct a process by looking only at the scoreboard.
Edge — local CNN vision at the oven exit, on 100% of the parts.
Total inspection is not solved by adding people: it is solved by putting the eye where cadence does not matter. Edge is a terminal with a local GPU that infers in milliseconds, trained on the parts and colors of your catalog, not on a generic laboratory model.
Cameras at the curing oven exit. The CNN infers per part in milliseconds: the defective one is pulled before it reaches assembly, and every defect is classified by type, rack position and paint lot, feeding root-cause analysis in real time.
How Edge works in paint inspection:
- Local, no cloud — inference happens on the line terminal. Neither the latency nor the cost of sending images to the cloud is compatible with paintshop cadence; and if the network drops, the line keeps inspecting.
- Trained on your catalog — the model learns from good and bad parts of your references and your colors. A high-gloss black and a matte grey do not fail the same way, and a generic model cannot tell what counts as a defect in your plant from what counts as finish.
- Classification, not just detection — "this part is bad" is not enough: the system distinguishes dirt inclusion, crater, popping, sag, orange peel or insufficient coverage. Without that detail, the root cause cannot be attacked.
- Every defect with its axes — type, rack position, color, basecoat lot and curing conditions. That is what turns a list of rejects into a map of causes.
- Pulled before assembly — the defective part leaves the flow at the point where it is still just a painted part, not an assembled module carrying added value.
Sampling inspection vs. 100% Edge inspection
| Aspect | Human sampling | With iLEAN Edge |
|---|---|---|
| Inspection coverage | A fraction of the parts | 100% of the parts on every rack |
| Moment of detection | At assembly or at the customer | At the curing oven exit |
| Cost of the rejected part | Carrying assembly added value | Only the painted part |
| Cause data | A global reject count | Type, position, color, lot and curing |
| Consistency through the shift | Depends on fatigue | The same on the first part and the last |
| Operation without network | n/a | Edge keeps inferring locally |
Impact estimate for your plant — to be validated with your own 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.
- High-volume automotive paintshop with sampling-based visual inspection at the oven exit and an FTQ managed today with the result rather than the causes.
- Edge pilot on the oven exit: cameras, a terminal with a local GPU and model training on good and bad parts from your catalog. First value expected within a few weeks.
- Scrap reduction in the order of 30% or more by closing the detection → cause loop, as a starting order of magnitude. Estimate to be validated against your current ratio.
- Indicative payback of 5 to 12 months, strongly dependent on the starting scrap ratio: the worse FTQ is today, the faster it pays for itself.
- The value that does not show up in month one: the map of causes. After a few weeks classifying defects by type, color and basecoat lot, patterns appear that no sampling inspection could ever have shown.
And the fair question from the paint manager
"What if the camera pulls good parts and sinks my yield?" — that is the right concern, which is why the threshold is calibrated on your parts and not set at the factory. Hallucination, moreover, is a problem of free generation: classifying an image against a pattern trained on your own catalog is an anchored task, where the best models brought the error below 1.5% [1]. And pulled parts are not scrapped automatically: they go through review, so the system learns from every discrepancy and the person keeps the final word.
[1] OpenAI paper "Why Language Models Hallucinate", 2025 — on the reliability of AI in anchored tasks.
What people ask about Edge inspection of paint defects
Does it keep up with the real cadence of a high-volume paintshop?
Yes, because inference happens locally, on a GPU terminal at the line side, and is counted in milliseconds per part. That is the reason for not using the cloud: neither the latency nor the cost of continuously uploading images is compatible with paintshop cadence. An important side effect of that decision is robustness: if the plant loses connectivity, Edge keeps inspecting and pulling, because its critical loop does not depend on the network. What syncs afterwards is the record, not the decision.
Why does the model have to be trained on our parts?
Because in paint, what counts as a defect depends on the product and the color. A high-gloss black shows a dirt inclusion that a textured grey hides; a sag in a non-visible area of the module may be tolerable and on a class-A surface may not. A generic model trained on laboratory images does not know those boundaries and produces two problems at once: it rejects the good and lets the bad through. By training on good and bad parts from your own catalog, the CNN learns where your real limit is — the one your customer argues about.
What happens to parts it pulls that were actually good?
They are not scrapped automatically: they go to a review lane where a person decides. That loop is not only a safety net, it is the learning mechanism: every discrepancy between what the CNN classified and what the person determined is used to tune the threshold. In the first weeks it is normal to calibrate towards caution and correct afterwards, because the cost of pulling a good part is recoverable and the cost of letting a bad one through is not. The final decision on the part's fate remains human.
Which defects does it distinguish exactly?
The ones inherent to the process: dirt inclusion, crater, popping, sag, orange peel and insufficient coverage, plus shade deviations when it is instrumented for that. What matters is not only detecting that a part is bad, but classifying what is bad about it: a dirt inclusion points to booth or air contamination, a crater to silicone or surface preparation, popping to curing. Without that distinction the data serves to count rejects but not to attack causes, which is where the real return lies.
How is the loop between detection and root cause closed?
By cross-referencing the defect with the variables that explain it — the ones other cases in the matrix have put into the system: the basecoat lot and the startup sheet conditions, the oven curve for that rack and the lab measurements for that shift. Every Edge inference is born linked to lot, color, rack and conditions, so asking "are the craters concentrated in one basecoat lot or in a humidity band?" becomes a query. That closed loop is what turns an inspection into an FTQ improvement tool.
Send us 200 photos of good and bad parts and we will show you what the CNN sees.
We work on your plant's real data, not ours. With your own parts we show you what it detects and with what margin. Assessment with no commitment.
See how we apply it in your plant — trial on your own parts ‹ See all 12 paint & assembly cases See automotive