Key points
- People are good at judgement, new faults and exceptions; cameras are good at the same defined check on every unit.
- The research literature treats human error in inspection as something that can be reduced but not eliminated.
- Checks at intervals miss faults that start and stop between two checks.
- Automate first the checks that are defined, visible, frequent and costly when missed: labels, codes, caps.
- Keep people for reviewing rejects, new products, judgement calls and fixing the cause.
What people do well
Experienced inspectors notice things nobody told them to look for: a new kind of fault, a smell, a pack that feels wrong. They judge borderline cases with context, adapt instantly to a new product, and, most importantly, find and fix the cause. No camera replaces that.
Where manual inspection struggles
A literature review by Sandia National Laboratories sums up decades of research: human inspectors are imperfect, and inspection error can be reduced with appropriate interventions but cannot be eliminated. Error rates of 20 % to 30 % are frequently quoted in the inspection literature across many kinds of inspection task, and even 100 % inspection does not find every defect (See, Visual Inspection: A Review of the Literature, Sandia 2012).
The same review reports that inspection accuracy suffers as defects become rarer, and that performance over long, repetitive watches tends to decline most for rare events, difficult detection tasks and tasks without feedback. It also notes a debate about how far those laboratory findings apply to real inspection, where how visible the defect is and how complex the decision is may matter most (See, Visual Inspection: A Review of the Literature, Sandia 2012). On a packing line, that describes a person checking labels for a mix-up that happens rarely, on every pack, for a whole shift.
Checks at intervals add a second gap: a coder that fails between the 10 o’clock and the 10:30 check prints bad codes for up to half an hour before anyone looks. The usual reports sound familiar: “label check every 30 minutes and at every reel change”, “sign off dozens of labels a day on top of the rest of my work”.
What a camera adds, and where it stops
A camera station checks every unit at line speed against the same limits, keeps counts and images for each batch, and stops or rejects on the first bad pack. It does not get tired, but it only finds what it was set up for: visible faults with limits agreed on real samples. It needs a recipe for each product, it produces some false rejects, and it cannot notice a new kind of fault the way a person can.
Side by side
| Aspect | Manual inspection | Camera inspection |
|---|---|---|
| Coverage | Samples at intervals, or 100 % at limited speed | Every unit at line speed |
| Consistency | Varies between people, shifts and over time | The same limits on every unit, until the recipe is changed |
| New or unusual faults | Strong: people notice what they were not told to look for | Weak: finds what it was set up and trained for |
| Judgement calls | Strong, with experience | Only within limits agreed in advance |
| Records | Sign-off sheets and check forms | Counts, results and images per batch |
| Changeovers | Quick to adapt; relies on training | Needs a recipe for each product; selected by the PLC |
| What it cannot see | Fast, small or rare faults; hidden features | Anything not visible to the camera; faults it was not set up for |
What to automate first
Automate the checks that are defined, visible, frequent and costly to miss. On most packaging lines that means labels after changeovers, date and batch codes, and caps and tamper bands. Leave judgement-based and non-visual checks with people.
| Check | Cost of a miss | Suited to a camera? | Suggestion |
|---|---|---|---|
| Wrong label or packaging after a changeover | High: recall and allergen risk | Defined and visible | Automate early |
| Missing, unreadable or wrong date and batch codes | High: legal and recall scope | Defined and visible | Automate early |
| Missing or crooked caps, missing tamper bands | Medium to high | Defined and visible | Automate early |
| Counts in open packs and trays | Medium | Defined when items are visible | Good candidate |
| Cosmetic surface marks with no clear limit | Varies | Needs agreed limits first | Agree limits, then decide |
| Taste, smell, texture, internal faults | Varies | Not visible to a camera | Keep other methods |
People and cameras together
- The camera checks every unit; people review the rejects. Rejected packs and their images tell the team what is going wrong and whether limits need attention.
- People set up new products. Teaching a new recipe on samples is a trained task for the site team or the integrator.
- People test the station. Challenge tests with marked test packs prove the camera and the reject still work.
- People fix the cause. The station reports the symptom; the setter, the coder technician or the supplier fixes it.
Skills and shifts in South African plants
South African packaging trade press describes a skills shortage in which many experienced operators, artisans, setters and supervisors are approaching retirement, and some of the skills plants need are not found locally (Packaging & Print Media, “Mind the gap”). Automating a repetitive visual check frees scarce experienced people for the work only they can do. It does not remove the need for them: someone still has to set up, test and look after the station.
How a station is chosen, tested and supported is described in how we deliver a project, and the checks themselves under inspection applications.