MANUFACTURING
Ceramic Defect Detection
Ceramic Defect Detection. What it takes to run it.
The chip and the pinhole boxed, the crack masked by area, on the tile line after the kiln, with a printed vein kept from reading as a defect.
// 01 · The question
“Is that line across the glaze a crack that costs a grade, or a printed vein on the stone range?”
// 02 · Why it is hard
A hairline crack is a few pixels wide on a glossy glaze, so the light has to throw it into a shadow line before any model can see it at all. A stone-effect range prints veins into the glaze that run across the face exactly the way a crack does, and a model shown only plain tile boxes every vein as a defect.
// 03 · What the labels have to be
Instance segmentation
Decided before the first frame
A chip breaks the tile's outline and a pinhole is a count, so a box carries both. A crack is graded by how much of the face it affects, and only a mask carries that area, so the extent-graded class sets the geometry the labeling runs under.
This is the decision that is expensive to reverse. The geometry has to match what the answer contains, and finding out it does not means labelling the set a second time.
// 03a · What it finds
An edge chip
boxed where it breaks the tile's outline, under the line's own light bank at the outfeed
A pinhole
boxed as a count on the face, since its size never sets the grade
A hairline crack
masked under raking light at a low angle, so the area it covers sets the grade the plant already grades by
A tile the model doubts
the stone-effect range, where a printed vein crosses the face the way a crack does, held for the inspector before anything trains on it
// 03b · What it checks
- 01
The crack's masked area against the plant's own grading tolerance, read off the mask rather than guessed from a box.
- 02
A vein on the patterned range against the plant's own labeled set of clean veined tiles, before a vein is ever written down as a crack.
- 03
The rule's schedule, cooldown and severity, approved before it goes live.
- 04
The inspector's override rate after a new glaze or a tightened tolerance, since neither looks like a failure on the day the frames stay the same.
// 03c · What the line provides
- The line camera already over the outfeed, under the raking light the plant already uses to throw a crack into a shadow line; no new camera.
- Frames from this line's own glazes, including the patterned ranges, not a plain-tile set from another plant.
- A labeled set of clean veined tiles from the patterned range, so the model learns the difference between a printed vein and a crack that breaks it.
- The plant's own grading tolerance in writing, since the model measures the crack's area and the tolerance is what turns that measurement into a grade.
// 04 · How we run it
Three parts of one loop, on this job.
LexAnnotate
Defect Annotation
Pixel-level defect annotation for surface scratches, cracks, dents, weld faults, and assembly misalignments. Multi-sensor alignment for multi-camera inspection systems.
LexInsight
Line Performance
Ask which shift or which line is producing the most of one defect class, and get an answer with the frames behind it, graded against the plant's own written standard.
LexAlert
Quality Alerts
Monitor defect detection accuracy across changing production conditions. Detect when lighting shifts, new materials, or equipment aging degrades model performance. Auto-retrain to maintain 99%+ accuracy.
// 05 · What breaks it after launch
A new glaze goes into production and the model has never seen a crack on it, or the grading tolerance tightens so a crack that was a second last quarter is a reject this quarter. The tile looks the same under the same light either way, and the first evidence is the inspector's override rate moving in one direction, not a drop any confidence score would show.
It is not the only one that can get this use case, it is the one that usually gets it first. All ten conditions.
// 06 · In manufacturing
Defect Detection at Scale
The Challenge
Vision models performed well in pilots but lost accuracy as factory conditions shifted.
Our Execution
Built pixel-level annotated datasets and multi-sensor alignment for reliable defect detection.
The Result
Maintained 99%+ defect detection accuracy across changing production environments. Reduced unplanned downtime.
99%+
Accuracy maintained in production
// 06a · Questions we get asked
Does this need a new camera on the tile line?
No. It runs on the line camera already over the outfeed, under the raking light the plant already uses to throw a crack into a shadow line. What it needs is frames from that camera on this plant's own glazes.
How does it tell a printed vein from a real crack?
Mostly from the plant's own frames. A stone-effect range prints veins that run across the face the way a crack does, and a model shown only plain tile boxes every one of them; a labeled set of veined tiles with no defects teaches it the difference. What is still unclear after that goes to the inspector before anything is scored.
Can it grade a crack by how much of the face it affects?
Yes, with a mask rather than a box. A box gives a crack's position; a mask gives its area, which is what a plant that grades by affected-face-area actually needs.
What happens when the plant changes glaze or tightens the tolerance?
The tiles look the same under the same light and nothing in the frame announces either change. The fix is a relabeled window of frames under the new glaze or the new tolerance; the signal if it goes unrecorded is the inspector's override rate climbing in one direction.
// 07 · Also in manufacturing