Most inspection systems that fail were decided at the lens, not at the model. Resolution, shutter type, optics geometry and how the camera learns a part has arrived are settled before any image is captured — and once wrong, no amount of training data recovers them.
The useful figure is pixels across the smallest feature that must be detected, not megapixels. A defect needs several pixels across it to be distinguishable from noise — three at absolute minimum, and five or more for anything that has to be measured rather than merely found.
So the calculation runs backwards from the defect: smallest feature size, times the pixels needed across it, divided into the field of view. That gives the sensor resolution. Choosing a camera first and hoping the defect resolves is the most common and most expensive sequence error in this category.
Field of view should be as tight as the part tolerances allow. Every millimetre of extra view spends resolution on space where a defect cannot occur, and part positioning variation is what forces the margin. Better fixturing frequently buys more than a bigger sensor.
| Decision | The options | What it turns on |
|---|---|---|
| Shutter | Global or rolling | Anything moving during exposure. A rolling shutter exposes rows sequentially, so a moving part is skewed. On a conveyor, global is not a preference |
| Scan type | Area or line scan | Continuous web or cylindrical parts suit line scan, which builds an image as the material moves and needs an encoder to stay square. Discrete parts suit area scan |
| Monochrome or colour | Mono, colour, or multispectral | Mono gives full resolution to every pixel and better low-light performance. Colour costs resolution to the filter array and is only worth it when colour is the defect |
| Interface | GigE Vision, USB3 Vision, CoaXPress, MIPI | Cable length, bandwidth and whether the camera is powered over the same cable. GigE reaches 100 m and is simplest to integrate; USB3 is faster over short runs |
GenICam gives a common programming interface across these transports, so the software layer need not change when the transport does.
The controller knows where parts are; the vision system does not. Position comes from a sensor or encoder into the PLC, and the PLC issues the trigger.
A digital input asserted by the PLC gives repeatable latency measured in microseconds. A software trigger over the network is subject to whatever the operating system is doing, and jitter becomes position error on a moving line.
The camera asserts a strobe output at the start of exposure so the illumination is synchronous with capture rather than merely nearby. On a moving part, that difference is blur.
The permitted blur is a fraction of a pixel at the part's speed. That sets maximum exposure, which sets required light, which sets the illumination design.
Pass or fail comes back to the PLC as a digital output or a fieldbus word within the station's cycle budget, because the PLC actuates the reject.
Image, measurements and parameters go to the MES or historian on a slower path, so record-keeping never sits inside the control loop.
Work backwards from the defect. Take the smallest feature that must be detected, multiply by the pixels needed across it — three at absolute minimum, five or more if it must be measured — and divide into the field of view. Megapixels alone tell you nothing without the field of view.
Whenever the part moves during exposure. A rolling shutter exposes rows sequentially, so a moving part is captured skewed. On a conveyor or an indexing line, global shutter is a requirement rather than a preference.
Keeping magnification constant regardless of distance, so a feature does not change size as the part sits higher or lower in the fixture. Essential for dimensional measurement; unnecessary and expensive when detecting surface defects.
From the PLC, as a hardware digital input. That gives repeatable latency in microseconds. A software trigger over the network inherits whatever the operating system is doing, and that jitter becomes position error on a moving part.
Area scan captures a two-dimensional image in one exposure and suits discrete parts. Line scan builds an image one row at a time as material moves past, which suits continuous web or cylindrical parts — and needs an encoder so the image stays geometrically correct when speed varies.
A common programming interface across camera transports such as GigE Vision, USB3 Vision and CoaXPress. It means the software layer does not have to change when the transport does, which matters when a camera is replaced years later.
Because deepening it means closing the aperture, which costs light, which lengthens exposure, which on a moving line causes motion blur. Depth of field, light and speed form the real optical design triangle.
Written for engineers. Share it with one.
The full inspection deployment.
SOLUTIONInspection at the station that made it.
KNOWLEDGEVerdict inside the cycle.
PLATFORMImaging chain and edge inference.
PRODUCTWhere the model runs.
INDUSTRYThe line will not slow down.
Tell us the specification, the constraint and the deadline. Programmes that cross silicon, radio, embedded and AI are where Faststream is strongest.