Surface defects, caught before packaging

The problem

Human visual control at line speed is an inspection by sampling.

100% human visual control is impossible at real line speed. Sampling inspection leaves gaps that end up as full-pallet returns.

• At kiln-exit speed, thousands of square meters pass in front of the sorting line every shift. The human eye tires and lets through tiles with a hairline crack, a glaze bubble, a pinhole or a slight shade variation.
• 100% human visual control is not possible at real line speed. What actually exists is a sample, and the quality of that sample changes with the hour of the shift. Night shifts and long runs of the same design are where it drops most.
• The gaps in that sample end up as claims and full-pallet returns, because a single visible defect on a laid floor discredits the whole batch. The installer stops laying and the claim covers the whole lot, not one tile.
• And on large formats and rectified pieces, every tile that goes out defective costs more than on a small format, in freight, in claims and in reputation with the distributor. A distributor who receives one bad pallet looks at the next ten more carefully.

How it fits the IRIS system

Edge — a local vision model at the kiln exit, every tile, in milliseconds.

Edge — local CNN vision, trained on the plant's actual format and collection, with no need to send images to the cloud.

The model is trained on your actual formats and collections, not on a generic catalog, and inference runs on the line. No image has to leave the plant for a tile to be classified before it reaches packaging. The sorting team stops looking at every tile and starts deciding on the doubtful ones. And the criterion no longer depends on who is on shift or how many hours they have been at it.

See the full IRIS architecture →

Before and after

Kiln-exit inspection today versus inspection with Edge

Aspect	Human inspection	With iLEAN Edge
Share of tiles inspected	A sample	100% of the pieces
Hairline cracks and pinholes	Missed when the eye tires	Classified tile by tile
Glaze bubbles and shade variation	Depends on the shift	Constant criterion
Point where a defective tile is caught	At the customer	Before packaging
Tile images leaving the plant	—	None: local inference
Doubtful tiles at sorting	Pass or fail by intuition	Escalated to a person

Partial sampling → 100% piece control. Visual-defect claims → drastic reduction.

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 scrap ratio. The higher your claims today, the shorter it is likely to be.
• From partial sampling to control of 100% of the pieces at kiln exit.
• A drastic reduction in visual-defect claims and in returns of full pallets.
• And each rejection is stored with its image, which shows which defect is growing and on which format, before it becomes a claim. That data is also useful upstream, at the press and on the glazing line, to remove the cause.

Estimated payback 5-12 months depending on current scrap ratio, reduced claims and returns. *Estimate to validate*.

And the fair question from the production manager

“What if it rejects good tiles because of the design?” — on decorated porcelain the graphic itself can look like a defect, which is why the model is trained on good and bad pieces of your own collections. Classifying a known defect on a known design is an anchored task, where the best models drop below 1.5% error [1]. And borderline pieces are not simply rejected: they are escalated to a person, and every decision feeds back into training. The threshold for rejecting is agreed with quality, not imposed by the model.

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