Skip to content
LexDataLexData
PlatformIndustriesCustomers
DocsThe Field GuideBlogWhy models drift
AboutCareersSecurityContact
Log inStart now

// MANUFACTURING

Foreign Material Detection

Foreign Material Detection. What it takes to run it.

The bone, the glove fragment and the seal wrinkle boxed on the fillet line's own camera, with the portion count from the same frame.

Start with your footageBook a demo →All of manufacturing →

// 01 · The question

“Which tray has the bone, the glove fragment or the short portion the twice-an-hour walk would have missed between passes?”

// 02 · Why it is hard

No two fillets are the same shape, colour or size, so the rare classes, a bone, a fragment, a wrinkle, are just as varied as they are uncommon, and a model trained on a supplier's studio photograph finds nothing on a wet blue belt. The same conveyor also reads as two different conveyors between the morning skylight and the afternoon LEDs.

// 03 · What the labels have to be

Bounding box

Decided before the first frame

A bone, a glove fragment and a wrinkled seal are each a yes-or-no call on one tray, found by where they sit, not how much area they cover. A box answers that; a mask would cost labeling time measuring an edge nobody doses or disposes against, and the portion count is the same box drawn once per portion rather than a second shape.

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 foreign object in the tray

boxed on the conveyor's own camera, a bone the trimmer missed, a glove fragment, a scale patch, among fillets that are never the same shape twice

A wrinkled or folded seal

the tray's film where it meets the flange, boxed on the same frame as the foreign-object check

The portion count

a box per portion in the tray, so a short tray is caught before it reaches the case rather than at a customer's table

A crop the model cannot place

a shadow that could be a bone, a skin patch that could be a scale patch, held for a person rather than scored either way

// 03b · What it checks

  1. 01

    The box against a person's eye, mostly at the line between a shadow and a bone or a skin patch and a scale patch, since irregular product has no two frames alike.

  2. 02

    The rule's own severity and cooldown: a foreign object set critical, the frame to the technician's phone so she can pull the tray by hand; a short portion set routine, logged per shift and per trimmer rather than paged.

  3. 03

    A frame the model doubts, held for a person with the crop rather than passed or rejected on a guess.

  4. 04

    The technician's override rate through the year, since the same camera reads differently under June's skylight and November's LEDs, and the correction rate is what shows the light moved, not any one frame.

// 03c · What the line provides

  • The camera already over the conveyor; no new camera, no new hardware on the line.
  • Frames from this line's own belt and shift, not a supplier's product photograph, since no two fillets are the same shape, colour or size.
  • The class names in the technician's own words, a bone, a glove fragment, a scale patch, a seal wrinkle, so a reject is called one thing rather than two.
  • Someone to read the alert, pull the tray by hand, and label a new variant's frames before its first real run.

// 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.

How the loop fits together →

// 05 · What breaks it after launch

Covariate shift

The season turned

→Covariate shift →

Nobody changes the line between June and November, which is exactly why nobody looks. The morning shift runs under bright skylight in summer and under LEDs with a low winter sun slanting across the belt by mid morning, throwing one fillet's shadow onto the next. The first signal is not in any one frame but in the technician's own corrections climbing on that shift as the sun drops through the year.

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

Full write-up ↓Every case study →

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 pull the tray with the foreign object in it off the line?

No. It boxes the object and sends the frame as a critical alert to the technician's phone, severity and cooldown set before it goes live. Pulling the tray is still her own hand, now with the crop already in front of her instead of a clipboard and a torch.

How does it tell a bone from a shadow on a fillet that has no fixed shape?

Not perfectly on its own. No two fillets are the same shape, colour or size, so the frames that teach the model are this line's own, across shifts and species. A crop it cannot place is held for a person rather than scored either way, and that frame becomes the label the next version learns from.

What happens when the light changes from the morning skylight to the afternoon LEDs?

Nothing in any one frame announces it. What shows up is the technician's corrections climbing on the shift where the light moved, and that override rate is what gets a window of frames labeled and folded into the next version, not any model metric.

What does the short-portion count get checked against?

The line's own threshold, not a complaint after the fact. Short trays are logged per shift and per trimmer, so a question like which trimmer short-packs on a given afternoon has frames behind the answer.

Read next

Foreign material detection on a food processing line, the full walk-through →Surface defect detection: the same rare-class imbalance on any part →What the skylight-to-LED light change does to an untrained model →
Start with your footageBook a demo →

// 07 · Also in manufacturing

Surface Defect Detection→Weld Quality Inspection→Assembly Verification→Package & Label Inspection→Date & Lot Code Inspection→Predictive Maintenance→Multi-Sensor Monitoring→Tablet Defect Inspection→Syringe Barrel Defect Detection→All forty-one→
Surface Defect DetectionWeld Quality InspectionAssembly VerificationPackage & Label InspectionDate & Lot Code InspectionPredictive MaintenanceMulti-Sensor MonitoringTablet Defect InspectionSyringe Barrel Defect DetectionAll forty-one →

Send us a week of this footage. We’ll show you what comes back.

Start with your footageBook a demo →
LexData
LexData

Product

  • Platform
  • Industries
  • Use cases

Resources

  • Docs
  • The Field Guide
  • Blog
  • Why models drift

Industries

  • Energy & utilities
  • Oil & gas
  • Agriculture
  • Manufacturing
  • Insurance
  • Retail
  • Robotics

Company

  • About
  • Customers
  • Careers
  • Contact

Trust

  • Security
  • Privacy
  • Terms

Stay updated

What we learn running vision models in production.

See everything.
Miss nothing.

Stay updated

What we learn running vision models in production.

Terms of use & Privacy policy

© 2026 LexData Labs · All rights reserved