Visual quality inspection on the line
Consistent inspection at line speed, on the defects you can actually show examples of.
Vision inspection works well on defects that look like themselves. It works badly as a general-purpose replacement for an experienced inspector.
Typical timeline: Feasibility 3–4 weeks, pilot 10–16 weeks
Imaging first. The model is the easy half.
These are the signs this is worth doing
If several of these are true, this is usually where the fastest return sits.
Vision inspection is genuinely capable now, and it is also the area where we most often advise against proceeding — not because the technology fails, but because the problem as described is frequently not the problem that exists.
The systems work well on defects with consistent visual signatures: surface marks, missing components, print and label errors, dimensional variance, fill levels, seal integrity. They work poorly as a general substitute for an inspector who is drawing on twenty years of noticing that something is subtly off.
What vision does well, and badly
Reliable
- Surface defects with consistent appearance
- Missing or misplaced components
- Label, print and barcode verification
- Dimensional checks against tolerance
- Fill level and seal integrity
- 100% coverage at line speed
Unreliable
- Defects you cannot show examples of
- Anything requiring judgement about acceptability
- Highly variable or reflective surfaces
- Novel defect types never seen before
- Inconsistent or changing lighting
- Subtle problems an inspector 'just knows'
Lighting and mounting decide the outcome
The most common cause of a disappointing vision project is not the model. It is that the images are inconsistent — ambient light changing through the day, vibration, parts presented at varying angles, a camera positioned where there happened to be room.
Controlled, repeatable imaging is most of the engineering, and it is unglamorous enough that it gets under-specified. We would rather spend the first phase on fixturing and lighting than on model tuning, because a good model on inconsistent images will never be reliable and a modest model on consistent images usually is.
How an honest pilot runs
- Pick one defect type with real examples — a few hundred images, including the ambiguous ones.
- Fix the imaging before touching the model: lighting, mounting, presentation.
- Run alongside human inspection, without acting on the output.
- Compare against inspector calls, and investigate every disagreement in both directions.
- Decide on measured false-accept and false-reject rates, not on a demonstration.
- Scale to a second defect type only after the first is trusted on the floor.
If you cannot produce a few hundred photographs of the defect, the honest answer is that it is not ready for a vision system yet.
We would rather tell you now than three weeks into a project. This work is usually the wrong call if any of the following describes you.
- Defects nobody can exemplify. Without examples there is nothing to learn from and nothing to validate against.
- Very low volumes, where an inspector is both cheaper and better.
- Environments where imaging genuinely cannot be controlled, and fixing that is out of scope.
- Anyone expecting to remove inspectors entirely. Coverage improves; judgement still belongs to people.
The systems involved
We integrate rather than replace wherever it makes sense. These are the systems this work most commonly touches.
Quality inspection — questions we get asked
How many images do we need?
For a well-defined defect, a few hundred good examples including borderline cases. Quality and coverage of the variation matter far more than raw count.
Will it replace our inspectors?
No, and we would be cautious of anyone claiming otherwise. It gives 100% coverage on defined checks and frees inspectors for judgement work and investigation, which is where they add most.
What about false rejects?
They are the real operational cost, and the trade-off against false accepts is yours to set. We measure both during the parallel run so the decision is made on your numbers.
The services this work sits inside
Where this comes up most
The sectors where we most often do this work, and where the payback is usually clearest.
Others worth reading
Delivery challans and weighbridge slips
Capturing proof of delivery at the gate, from a phone, closes the loop between what left and what was invoiced.
Read itVision AIInvoice data extraction and posting
Extraction is mature enough to post most invoices automatically. The design question is what happens to the ones it is unsure about.
Read itThinking about quality inspection?
Start with a short conversation. We will tell you honestly whether this is the right place to begin, or whether something else pays back faster.