MANUFACTURING
Glass Defect Detection
Glass Defect Detection. What it takes to run it.
The scratch and the rim chip boxed on the frame each one needs, so a reflection stops reading as a defect.
// 01 · The question
“Is that bright line on the bottle a scratch, a chip, or the ring light again?”
// 02 · Why it is hard
Glass does not hide its background the way a part does, so a scratch, a mould seam and a reflected ring light arrive on the same bright-line pixels unless the station's light is built to separate them, and a rim chip a millimetre wide and a hairline the height of the bottle are two different scales no single camera resolves at once.
// 03 · What the labels have to be
Bounding box
Decided before the first frame
A scratch, a rim chip and a hairline are each a yes-or-no call at a known place on the bottle, found by where they sit under the right light, not by how much area they cover. A box answers that; the harder part is building the frame the box gets drawn on.
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
A scratch on the bottle's body
the low-angle lit frame, where a true scratch scatters light back at the lens and a mould seam or a reflection does not
A chip or a hairline at the rim
the backlit rim frame, a chip read as a bite out of a crisp silhouette, a hairline as a faint fracture across it
A seed, a stone or a cord from the melt
either frame, boxed and classed in the glass trade's own words rather than one generic defect label
A mark the model doubts
a scratch tangled with a seam, or a reflection that survived the low-angle light, held for a person before anything trains on it
// 03b · What it checks
- 01
Every box against the frame it belongs to: rim defects read on the backlit frame, body defects on the low-angle frame, never the other way round.
- 02
The class against the glass trade's own words, seed, stone, cord, check, scratch, chip, rather than one generic defect box.
- 03
The rule's schedule, cooldown and severity, approved before it goes live.
- 04
The inspector's override rate on a new container spec, since a mould change dates every label without changing a single pixel on the station itself.
// 03c · What the line provides
- The two station cameras the line already runs for this inspection, one backlit on the rim, one low-angle on the body; no new camera.
- A plain matte backdrop and a short enclosure, so the conveyor and the hall's own skylight stop showing through the glass.
- Frames from this station's own lights and this container's own mould, not a bench or phone photograph.
- A labeled window of the new container before a mould change ships, since the old labels describe a bottle that is no longer running.
// 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 lighter mould changes the wall thickness and the shoulder profile, so the reflections sit in new places and the rim's silhouette reads differently, while nothing on the station itself changed. The first evidence is the inspector overriding the model in one direction on the new bottle, 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 bottle line?
No. It is built on the two cameras the station already runs for this inspection, one backlit for the rim, one low-angle for the body. What it needs is frames from those two cameras under this station's own light.
How does it tell a scratch from a reflection?
Mostly from the light, not the model. Under a low-angle light a true scratch scatters light back at the lens and a reflection or a mould seam does not, so the two stop looking the same once the frame is built that way. What still looks doubtful after that goes to a person before it is scored either way.
Can it catch a hairline crack the same way it catches a rim chip?
Only if each gets its own camera. A chip is a millimetre or two at a known place; a hairline can run the height of the bottle at a fraction of a pixel wide. One camera framed for the whole bottle resolves neither well, which is why the rim and the body each get their own frame and their own question.
What happens when the customer changes the bottle's mould?
The old labels describe a container that is no longer running. A window of frames from the new mould is labeled before the changeover where possible, and the inspector's own override rate on the new bottle is the first sign the old model is now wrong in one direction, not a score any dashboard would show.
// 07 · Also in manufacturing