MANUFACTURING
Strip Surface Defect Inspection
Strip Surface Defect Inspection. What it takes to run it.
The scratch, the inclusion and the scale boxed on every metre of the coil, the doubtful metres held for the inspector in the pulpit.
// 01 · The question
“Is that a scratch that costs ten metres of coil, or an oil mark that costs nothing?”
// 02 · Why it is hard
An oil mark and an inclusion throw back the same few pixels of contrast to a model that has seen only a few hundred of each, the class names have to survive a change of inspector, clean metres vastly outnumber defective ones by construction, and a class like crazing appears on a handful of coils a year, so a training set sampled at random loses it entirely.
// 03 · What the labels have to be
Bounding box
Decided before the first frame
Disposition is read per metre off the line's own encoder, and a scratch's length along the strip is the length written on the coil card. A box's length already carries that measurement; a mask would answer a question about area that nobody at the pulpit is asking.
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 along the strip
boxed at line speed on the finishing-line camera, the box's own length read as the defect's length against the coil card
An inclusion, a patch of crazing or rolled-in scale
boxed and classed in the pulpit's own words, since the sheet's names have to survive a change of inspector
The strip position for every box
read from the line's own encoder at the moment of the frame, so a metre's disposition is tied to where it sits on the coil rather than guessed from the frame alone
A metre the model doubts
between the pass and fail thresholds on any class, held in the review tier with the frame attached for the inspector in the pulpit, on nights as well as days
// 03b · What it checks
- 01
Each metre's class and length against the pair of thresholds that sets pass, review or fail, so a scratch across ten metres and a coil-length run of scale are graded by their own numbers rather than one pass or fail line.
- 02
The review queue on the day shift and at the morning clear on nights, since the pulpit is unmanned overnight and the queue is what keeps a doubtful metre from being guessed rather than held.
- 03
The rare classes on their own recall figure rather than folded into one accuracy number the clean strip would dominate, so a class like crazing is checked on the handful of coils it actually appears on.
- 04
The inspector's correction rate by class after a work-roll change or a new steel grade, since the defect mix moving is the first sign a threshold tuned on the old mix is now wrong.
// 03c · What the line provides
- The camera already over the finishing line, under the same bank of lights the pulpit inspector reads today; no new camera.
- Frames from this line's own footage, not a general steel-defect set, since the four classes are typed in this mill's own words and have to survive a change of inspector.
- Every frame of a rare class the line has ever recorded, kept and weighted, since a handful of crazing frames a year is lost entirely in a training set sampled at random.
- The line's own encoder reading on each frame, so a box carries which metre of the coil it belongs to.
// 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 set of work rolls or a steel grade the pickle line is still learning changes the scratch and inclusion counts while the strip itself looks the same under the same lights. Every threshold tuned around the old mix is now wrong, the fail tier floods with one class or the review queue fills with a class that turns out to be nothing, and the first sign is the inspector's correction rate on that class, 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 replace the inspector in the pulpit?
No. It boxes every defect on every metre at line speed, day and night, and sorts each metre into pass, review or fail. The review tier is where the inspector's call still counts, on the metres the model sits between its own thresholds on.
How does it tell a scratch from an oil mark?
Mostly from the inspector's own corrections. An oil mark and an inclusion look alike to a model that has seen only a few hundred of each, and every oil mark the inspector un-boxes teaches the next version to leave oil marks alone, the same discipline as any other review tier.
What happens on the night shift when the pulpit is empty?
The review queue still fills, with the frame and the box attached, and it waits. The day inspector clears it the next morning before the night's coils are dispatched, so a doubtful metre is held a few hours longer rather than guessed.
What happens when the mill changes the work rolls or the steel grade?
The per-metre read does not move, but the counts do: the scratch rate can fall while the inclusion rate climbs, and a threshold tuned around the old mix starts failing metres that should pass and filling review with metres that should not be there. The first sign is the inspector's correction rate on one class, which is why the thresholds are re-tuned against the new mix before anyone retrains.
// 07 · Also in manufacturing