VISUAL QUALITY INSPECTION
Human inspection is accurate for about twenty minutes
Visual checks are consistent at the start of a shift and drift after that, which is exactly when the sampling plan says a defect should be caught. Automated inspection does not get tired, but it only pays back when it is tied to the rest of the plant: an image of a defect is a curiosity, whereas an image linked to the machine, the coil, the die change and the operator is a root cause. The second one is what stops the defect recurring.
What the system does
Defect taxonomy before cameras
We start by agreeing with your quality team exactly which defect classes matter and what distinguishes a reject from a concession. That classification work drives every later decision about optics, lighting and model.
Optics and lighting are the hard part
Most failed inspection projects failed at the lighting rig, not the model. Surface defects on drawn wire, plated finish and machined parts each need different illumination. We specify and trial this before committing to a model architecture.
Models that fit the sample you actually have
Defects are rare, so defect images are rare. Where there are not enough examples for supervised classification, we use anomaly-detection approaches trained on good product, which needs far less labelled defect data to start producing value.
Linked to batch and machine
Inline inspection at line speed where the cycle allows, offline station inspection where it does not. We will say plainly which is feasible for your line rather than promising inline and discovering the frame rate later.
Operator in the loop
Borderline calls route to an operator whose decision retrains the model. The system gets better because your people use it, rather than degrading quietly as product mix shifts.
How a visual inspection project runs
Defect study and sample collection
We collect physical samples of each defect class and good product. This is the single biggest determinant of whether the project succeeds, and it is work that has to happen on your floor with your quality people.
Optical bench trial
Camera, lens and lighting combinations tested against the real samples before any capital commitment. The deliverable is evidence that the defects are visible to a camera — or an honest finding that some of them are not.
Model development
Training against the collected dataset, with classification thresholds set jointly with quality. We tune against the cost of a missed defect versus a false reject, which is a business decision, not a technical one.
Line integration
Trigger from the PLC, result back to the PLC for reject actuation, images and verdicts into the MES. Runs in shadow mode against human inspection until the agreed confusion matrix is met.
What this is built on
Industrial cameras over GigE Vision with controlled lighting, triggered from the PLC so capture is deterministic rather than timed. Convolutional and anomaly-detection models in PyTorch, exported for inference on an edge machine next to the line so the verdict is available within the cycle. Images, verdicts and model versions stored with the production record in the MES database, so a defect found at the customer can be traced back to the exact frame and the exact model build that passed it.
Send us samples, not a specification
The fastest way to find out whether automated inspection is viable for your product is for us to look at real defective parts. Tell us what you make and what goes wrong with it, and we will tell you what a bench trial would involve.