MANUFACTURING
Tablet Defect Inspection
Tablet Defect Inspection. What it takes to run it.
Every tablet boxed and each reject classed as capping, lamination, a chip or a crack, so the class names the press setting to check.
// 01 · The question
“Which press setting made this reject, a chip, a crack, lamination or capping, named with the count instead of a tray to sort by hand?”
// 02 · Why it is hard
Capping and lamination look alike on a crop and point to different press settings, so the taxonomy has to come from the process engineer's own reference crops before anyone labels. A new formulation's first shift has no labelled crops for its own fault at all, and the dust and belt motion of a compression line are already on every frame.
// 03 · What the labels have to be
Bounding box
Decided before the first frame
The first question is only whether a tablet is there, which is a box, tuned to miss none even where it means catching a fragment or a shadow, since a tablet that never gets a box never reaches the second step. The fault itself is named by a tag on that box's own crop, not a second shape, because the question is which fault, not how much of the tablet's surface it covers.
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
Every tablet on the belt
boxed on the compression line's own camera, tuned to miss none even where it means a box catches a fragment or a shadow
The fault on a rejected tablet
a tag on that box's own crop, capping, lamination, a chip, a crack or contamination, named from the engineer's own reference crop for each
A fault the taxonomy has no class for yet
a new formulation's first crops, held for the engineer to name before anything is labelled against it
A crop the classifier cannot place
held for a person rather than forced to a class, the same default as any other low-confidence frame
// 03b · What it checks
- 01
The tablet box against a person's eye, mostly where two tablets touch and one box might be catching both.
- 02
The crop's class against the engineer's own reference crop, since capping and lamination can look alike and name different press settings.
- 03
A frame with more capped or cracked tablets than the rule's own count, held for the engineer with the crops attached rather than a bare number.
- 04
A crop the classifier is unsure of, held for a person before anything trains on it, the same review step as any other class.
// 03c · What the line provides
- The camera already over the compression line; no new camera, no new hardware.
- Frames from this press, not a reference photograph, since the classifier learns this line's dust and this belt's own lighting.
- The taxonomy in the process engineer's own words, with a reference crop per class, written before anyone labels.
- Someone to name a new fault class when one appears and to read the reject tray the count gets checked against.
// 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 formulation arrives with a fault the taxonomy has never named, and the crop shape is the same crop shape it always was. The finder still boxes every tablet; the classifier calls the new fault its nearest existing class until the engineer names one and a few dozen crops are labelled against it.
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 the camera tell me why a tablet failed?
It names the fault with a tag on the crop, capping, lamination, a chip, a crack or contamination, and each class points at a different press setting the engineer already knows to check. The camera finds and classes the crop; which setting actually caused it is still the engineer's read.
What happens on the first shift of a new formulation, before the taxonomy has a class for the fault?
The finder still boxes every tablet; a fault with no class yet gets called its nearest existing one, the usual sign being overrides climbing on that shift. The engineer names the new class, a few dozen crops are labelled against it, and a retrained version replaces the old one with no downtime.
Capping and lamination look alike, how does the system tell them apart?
Not perfectly on its own. The classifier is trained on this line's own confirmed crops, and capping against lamination is the usual confusion point even for it, which is why a crop it cannot place goes to a person rather than being guessed.
What does the process engineer receive?
A count held against the rule's own threshold, with the crops attached, sent to the shift rather than a dashboard. Twenty capped crops from the last hour is something to check on the press; a bare reject count is not.