Machine vision in production
A camera that sees what goes wrong the moment it happens. We start small to prove it works, then build the setup that can keep up with line speed.
The first thing we work out is not which model to use, but whether the defect can be made visible at all. That is a question of light and optics, not of AI.
Rapid prototyping: start with a trial, not with a quote
Before any hardware goes up in the hall, you want to know whether the defect is visible and whether a model finds it reliably. That can be done in a few days with a phone and a temporary setup. You collect real images of your own product, train on them, and see what comes out. If it turns out not to work, you learned that for the price of a week's work instead of an installation.
If it does work, you immediately know what the production setup has to handle: how sharp, how fast, and with what lighting.
What a phone really does, and where it stops
A recent iPhone has a 48 megapixel sensor and runs models on the device itself. For a trial that is enough, and for some manual inspection stations it stays enough. There are three limits you need to know before you build on it.
Temperature
Apple specifies 0 to 35 degrees Celsius as the operating range. Above that the device shuts down functions on its own to cool off. In a hall with process heat that is exactly the behaviour you do not want during continuous inspection.
Rolling shutter
The sensor reads out row by row, not all at once. With a product moving past, that produces distortion: the time difference between the top and bottom image rows shows up as skew.
No trigger and no fixed optics
You cannot start the capture from a hardware signal, and you cannot fit a telecentric lens. Without the first you are not certain which product you photographed, without the second you cannot measure reliably.
Where it is genuinely strong
- Rapid prototyping: a working trial in days rather than months
- Collecting first images to train a model on
- Manual inspection stations where someone holds the device
- Building a visual record for rejects or complaints
- 3D measurement with the LiDAR scanner, down to millimetres within five metres
The setup that keeps up with the line
Once it moves to production the question changes. No longer whether the defect is visible, but whether you catch it every time, at line speed, for years, in a cabinet that can take the hall.
Industrial camera
A global shutter exposes all pixels at once, so a moving product stands still in the image. You pick the interface on distance and bandwidth: GigE Vision reaches up to 5 Gbit/s over a hundred metres of cable, USB3 Vision is faster over short runs, CoaXPress goes up to 12.5 Gbit/s per channel.
Light and optics
This decides whether it works, more than the model does. Coaxial lighting along the optical axis gives even illumination without cast shadows. A telecentric lens keeps magnification constant, and that is the precondition for taking measurements from an image.
AI computer in the cabinet
A passively cooled industrial computer with an AI accelerator, on DIN rail in the control cabinet. Depending on what the model needs you are between 67 and 275 TOPS at 7 to 60 watts. No fan means no filter to clog and no part that fails first.
Back to the control system
The verdict has to do something. As a digital output to the PLC for ejecting or sorting, or as a variable in the control system's OPC UA model so the rest of the factory sees it too. Usually both.
We do this with partners
Optics, lighting and camera selection are a trade of their own. We build the software, the integration and the link to your control system, and bring in the supplier who picks the right lens and the right light. So you are not buying a piece of software on its own, and not a sealed package either, but a working setup you can get into yourself.
Sources
The technical claims above, with the source behind them.
- Apple iPhone bedrijfstemperatuur · 0 tot 35 graden Celsius
- Basler: rolling en global shutter · Hoe de uitleesvolgorde vervorming geeft
- Basler: vergelijking camera-interfaces · GigE Vision, USB3 Vision, CoaXPress
- Cognex: telecentrische lenzen · Constante vergroting voor dimensionele meting
- Cognex: verlichting voor machine vision · Coaxiaal, bright field en dark field
- NVIDIA Jetson Orin · TOPS en vermogen per module
- Apple Developer: Vision en Core ML · Modellen op het toestel zelf uitvoeren
- Nauwkeurigheid iPhone LiDAR · RMSE 4,89 mm binnen vijf meter, peer reviewed
Our approach
From advice to management, with one partner
The same rhythm for every project.
Advice
We start with a conversation, not with code. First clarity on what you need and what you do not.
Build
We build and integrate to measure, with technology that fits you. You own your data and your code.
Manage
We stay involved: monitoring, adjusting and growing with your business.
Shall we first see whether it can be made visible?
One conversation about your product and your line, and we know whether a trial is worth it.
Call 085 083 5775A 30-minute intake