A pixel has no size. It is an array index. It acquires a size only when you tell the software how much of the real world landed on it — and that statement is the single most consequential thing you will do in an image measurement session.
The number itself
Calibration in an imaging system reduces to one division:
scale = known length ÷ pixel span
Image something whose length you already know. Measure how many pixels it spans. Divide. If a stage micrometer's 1.000 mm division covers 500 pixels, your scale is 0.002 mm per pixel, or 2 µm per pixel.
From then on, every measurement is a multiplication. A feature spanning 312 pixels is reported as 0.624 mm. The software is not measuring your part. It is counting pixels and trusting your division.
What that number actually claims
This is where most calibrations quietly become dishonest. "0.002 mm per pixel" is not a property of the camera. It is a claim about an entire optical chain at one moment:
- this sensor, with this pixel pitch
- this lens or objective, at this magnification
- this working distance between lens and part
- this region of the field of view
- this plane — the plane the calibration artefact was sitting in
Change any one of them and the number is no longer a measurement of anything. The most commonly overlooked item is the last. A calibration is valid in the plane you calibrated in. If your artefact sat on the stage and your part sits on a 10 mm fixture, the part is imaged at a different magnification and every reading is biased — consistently, invisibly, and in the same direction.
Calibrate in the plane you will measure in, at the magnification you will measure at. If the part height changes, the calibration changes.
Deriving it properly
Use a traceable artefact
A steel rule is not a standard. Neither is a printed target, a coin, or a drill bit you measured with callipers. Use a stage micrometer, graticule or gauge block that carries a calibration certificate leading back to a national standard. Without that, the chain stops at your bench and an auditor will say so.
Span as much of the field as you can
This is the single biggest improvement most people can make. Suppose you can locate an edge to within about one pixel. Calibrate across 50 pixels and your scale carries roughly 2% uncertainty. Calibrate across 1,500 pixels and the same one-pixel ambiguity is under 0.07%.
relative error ≈ edge uncertainty ÷ pixel span
That error does not stay where you left it. It multiplies every subsequent measurement. A 2% scale error turns a 40.00 mm feature into 40.80 mm, which will fail a ±0.25 mm tolerance for reasons that have nothing to do with the part.
Use several divisions, not one
If your graticule has 10 µm divisions, do not measure one of them. Measure across a hundred and divide by a hundred. You get the same scale with a tenth of the sensitivity to how well you placed the two endpoints.
Focus first, and focus on the artefact
A defocused edge spreads over several pixels, and where you click within that spread becomes arbitrary. Focus, then calibrate, then do not touch the focus.
Five ways the figure quietly goes wrong
- Somebody changed the zoom. On a continuous-zoom stereo microscope, the detented positions are repeatable and everything between them is not. If your zoom is not at a detent, your calibration belongs to a position nobody can return to.
- The part is not in the calibration plane. Covered above, and worth checking first whenever readings are consistently a few percent out.
- The calibration was taken in the centre and the part is at the edge. Real lenses have distortion. It is smallest on axis and grows towards the corners. A scale derived from the middle of the field will under-report near the edges — typically by a few tenths of a percent on a decent telecentric lens, and by several percent on a cheap wide-angle one.
- X and Y were assumed equal. They frequently are not, and a single scale factor then mis-measures every diagonal. This has its own guide.
- Nobody wrote it down. A scale factor with no record of which artefact, which date, which objective and which operator is not evidence. It is a setting.
How Measuret handles it
Calibration in Measuret is a named list rather than a single field. Each row carries a name, a Scale X, a Scale Y and a unit — micrometres, millimetres, centimetres, metres or inches. A machine with several objectives keeps one row per objective, so switching optics means selecting a row rather than re-deriving a number while a part waits.
You can enter the scale directly, or derive it in place: measure a known span on the image, state what it really is, and Measuret does the division for that axis. Because X and Y are separate rows in the same table, calibrating both is the default rather than an advanced option.
Internally the calibration is a transform, and Measuret refuses to accept one that is numerically unstable — it will reject the calibration rather than return a figure it cannot stand behind. It also checks that converting a length into physical units and back again returns what it started with, so a calibration that silently loses precision at your working scale is caught rather than tolerated.
Every measurement is then held in the source image's own coordinates and converted through that transform. Zooming, panning, resizing the window or exporting a report cannot change a value, because none of them touch the numbers the measurement is made of.
See how calibration works in the application, including a calculator that performs the same division.
This guide covers the scale factor itself. Uncertainty budgets, repeatability studies and gauge R&R are a larger subject; if you need a defensible uncertainty statement rather than a defensible scale factor, talk to a calibration laboratory.