IMAGE RECOGNITION
Is the right thing in the right place?
It is the question behind most of what gets checked by hand — the right part at the station, the right SKU in the facing, the right pallet on the right dock. A camera can ask it of every unit instead of every tenth one, and keep the picture that proves the answer.
The newest line of our work. The models and the edge deployment are ours; there is no reference installation to name yet, and the honest way in is a paid pilot on your own images.
WHAT IT CHECKS
Six questions a camera can answer at line speed
None of these need a person to be looking at the moment it happens. All of them cost you when nobody was.
Is this the right item?
The unit in front of the camera is matched against what the system says should be there — by shape, print, colour and code together, not by barcode alone. A barcode confirms what the label claims; this confirms what the label is stuck to.
Is it in the right place?
Bin, slot, shelf facing, staging lane, station. The check is against your own location data, so a mis-slot is caught when it is cheap to fix rather than at the pick that fails three weeks later.
Is all of it there?
Case counts on a pallet, parts in a kit, items in a carton before it closes. Counting is the thing cameras are unreasonably good at and people are reliably bad at after the second hour.
Is the label right and readable?
Present, legible, the correct language and revision, and actually on the correct face. Reprints and rejected deliveries usually trace back to one of those four.
Is it the right size or condition?
Dimensions without touching, plus the damage and seal checks that decide whether something is accepted at goods-in — with the image attached to the decision.
Can you prove it later?
Every verdict keeps its image, its confidence and its timestamp against the unit, the location and the shift. When a customer calls about one unit in October, you look it up instead of apologising.
WHAT IT LOOKS LIKE
A verification log you can actually read
An illustration of the check log — what was expected at each location, what the camera saw, and how sure it was. Your version carries your locations, your items and your language.
| Location | Expected | Seen | Result | Confidence |
|---|---|---|---|---|
| A-14-03 | Filter cartridge 40µm | Filter cartridge 40µm | match | 99.2% |
| A-14-04 | Sealing ring 22mm | Sealing ring 25mm | wrong item | 97.8% |
| B-02-11 | Label DE/FR rev. 4 | obscured | ask a human | 61.4% |
| B-02-12 | Carton 600×400 | Carton 600×400 | match | 99.6% |
Illustrative layout. Not a screenshot of a customer system.
HONESTLY, THOUGH
When a sensor off the shelf is the right answer
Vision is sold as though it always needs a project. Often it does not, and we would rather tell you that early than quote you for one.
Buy a standard vision sensor when
- The part arrives in the same position every time, under the same light
- You are reading a barcode, a date code or a character string, nothing more
- One known feature decides it — present or absent, right colour or wrong
- The line is one line, and it is not going to change next year
- Your machine builder already supports the sensor and will keep supporting it
Build something of your own when
- There are hundreds or thousands of items and new ones arrive every month
- Presentation varies — angle, lighting, packaging, partial occlusion
- The decision combines several signals, not one threshold
- The verdict has to land in your ERP or WMS to be worth anything
- You need the evidence trail afterwards, not just a pass or a fail at the time
We would rather lose a vision project at the pilot than win it on a demo. The pilot runs on your images, from your line, under your lighting — and if the number that comes back does not beat what your people already achieve, we will say so and you will have paid for an answer rather than a system.
HOW IT GETS BUILT
Images first, promises later
01
Collect the images
From the actual location, under the actual light, including the awkward ones — the half-covered label, the pallet stacked wrong, the part at the odd angle. A model trained on tidy images fails on a real Tuesday.
02
Pilot with a number attached
We train on your images and report accuracy per class, including what it gets wrong and how confidently. That number, not a demo, is what tells you whether to go further.
03
Wire it into the work
The verdict goes where the decision is made — a light at the station, a hold in the WMS, a line on the goods-in report. A camera that only fills a dashboard has not changed anything.
STRAIGHT ANSWERS
What people ask before they start
How many images do you need to get started?
Fewer than most people expect for a first read — a few hundred per class is usually enough to tell whether the problem is tractable. Production accuracy needs more, and specifically needs the hard cases, which is why the pilot collects them rather than assuming them.
What accuracy is realistic?
For a clean identity or presence check, high nineties is normal. For subtle condition judgements it is lower and sometimes not worth doing. We report the number on your own images before you commit, because the honest answer varies more by problem than by vendor.
Do we need special cameras and lighting?
Usually an industrial camera and controlled lighting, yes — and that is the part people underestimate. Consistent lighting does more for accuracy than a better model does. Where a phone or an existing CCTV feed is genuinely enough, we will say so.
Does it run in the cloud or on our site?
On site, at the line, for anything that has to keep a takt time or survive a network outage. Training happens off-site; the decision does not need to leave your building, and for most operations it should not.
What happens when the model is not sure?
It says so. Every verdict carries a confidence, and below your threshold it routes to a person instead of guessing. A system that quietly guesses is worse than no system, because you stop checking it.
Bring us a location and a thing that keeps going wrong.
A short call is usually enough to tell whether a camera is the right answer here — or whether it is a process problem wearing a vision costume.