MANUFACTURING
Contact Lens Defect Inspection
Contact Lens Defect Inspection. What it takes to run it.
Cracks, bubbles and edge tears boxed on the dark-field frame, the band between two thresholds held for a person before pass or fail.
// 01 · The question
“Is that bright line on the lens a crack, or the saline meniscus again?”
// 02 · Why it is hard
A clean lens is nearly invisible under ordinary light, which is why the station lights it from the side instead. The saline meniscus at the cup wall catches that same light and throws a bright arc that a first-week labeler boxes as an edge tear, so the hard part is a labeling rule, not a hard lens.
// 03 · What the labels have to be
Bounding box
Decided before the first frame
A crack, a bubble or an edge tear is a bright feature's presence and strength against two thresholds, not a measured area; the quality engineer's two thresholds decide the bucket, not how much of the lens a mask would cover. A box finds the feature and ranks it; a mask would cost labeling time on an extent nothing downstream reads.
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 crack across the lens
a bright line boxed on the dark-field frame, where a clean lens bends the light away and only a defect scatters it back at the camera
A bubble in the polymer
a bright ring on the same frame, boxed and classed apart from a crack even though both are bright features on the same black field
An edge tear or inclusion
a bright notch in the lens's faint boundary circle, told apart from the saline meniscus arc that catches the ring light at the cup wall
A lens the model cannot decide
a feature that lands between the two thresholds, held for a person with the frame and the proposed class before the lens is scored either way
// 03b · What it checks
- 01
Every feature against the two thresholds the quality engineer set, so noise, the band and a defect are three answers rather than a call forced into two.
- 02
The bright arc at the cup wall against the labeling rule that it is never a defect, so the saline meniscus never trains as an edge tear.
- 03
The rule's own severity and cooldown, a crack set critical on a medical line, approved before it goes live.
- 04
The review queue against the correction rate rather than the raw doubted count, since a wider band from a new polymer lot is a threshold question and a rising correction rate is a retrain question.
// 03c · What the line provides
- The dark-field camera and ring light already at the station; no new camera, no new hardware.
- Frames from this cup mount, under this line's own dark-field ring, with the saline meniscus in every one.
- The two thresholds in the quality engineer's own words, set from what a false reject costs against what a shipped crack costs.
- Someone to clear the band, the way the inspector already tilts the cup against black and decides.
// 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 polymer lot arrives and the bubble rate doubles while the crack rate holds. The lenses look the same and the classes mean the same thing; how often bubbles occur has moved, so the two thresholds tuned to the old rate now send far more lenses to the band than before, and no retrain shrinks that queue.
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 replace the inspector checking the tilted cup?
No. It puts a box on every bright feature the dark-field camera catches and sorts each lens by the engineer's two thresholds. A lens in the band still goes to a person with the frame, the same decision the inspector already makes by tilting the cup against black.
How does it tell a crack from the saline meniscus?
By a labeling rule written before training starts: the bright arc the ring light throws at the cup wall is never a defect, however closely it reads like an edge tear. Every labeler holds to that rule, so the meniscus never becomes a class on its own.
What happens to a lens between the two thresholds?
It goes to a person with the frame and the model's proposed class. The band is the point: a two-answer gate would force a call on weak evidence, and the weak calls are where a shipped crack comes from.
What happens when a new polymer lot changes the bubble rate?
The pixels do not move, so the model is not wrong; the thresholds tuned to the old rate now send more lenses to the band than the shift can clear. The review queue is the signal, a jump in doubted lenses with no jump in corrections, and the fix is a threshold change before anyone considers a retrain.
// 07 · Also in manufacturing