Almost every introduction to image measurement quietly assumes that a pixel is square in object space — that moving one pixel right covers the same real distance as moving one pixel down. It is a convenient assumption, it is often nearly true, and when it is not true it fails in a way that is very hard to spot from the readings alone.
The square-pixel assumption
With a single scale factor s, the distance between two points is
d = s × √(Δx² + Δy²)
That is only correct if one pixel of Δx and one pixel of Δy represent the same physical length. When they do not, the correct expression needs two scales:
d = √((sxΔx)² + (syΔy)²)
Notice what this does to shapes. A perfectly round bore imaged through an anisotropic system does not land on the sensor as a circle. It lands as an ellipse, stretched along whichever axis has the finer scale. Measure its "diameter" horizontally and vertically and you get two different answers — and both of them are, in pixels, correct.
Where anisotropy actually comes from
It is more common than the tidy diagrams suggest.
- Non-square sensor pixels. Less common on modern industrial sensors than it was, but far from extinct, particularly on older and on video-standard cameras.
- Analogue capture and resampling. Anything that digitises a video signal line by line samples along the line at a rate unrelated to the line spacing. Aspect ratios like 1.09 : 1 were a normal consequence.
- Software that resizes without preserving aspect. An image stretched to fill a fixed frame arrives at the measurement stage already anisotropic.
- Anamorphic or cylindrical optics. Rare in inspection, decisive when present.
- The part not being perpendicular to the optical axis. A part tilted about one axis is foreshortened along that axis only. This one is not really a pixel problem — it is a fixturing problem wearing a pixel problem's clothes — but it presents identically, and an anisotropic calibration will paper over it while the true geometry stays wrong.
How large is the error?
Take a modest 3% difference between axes: sx = 2.00 µm/px, sy = 2.06 µm/px, and suppose someone calibrated on the horizontal only and applied 2.00 to everything.
| Measurement direction | Error |
|---|---|
| Horizontal | 0% — correct by construction |
| 45° diagonal | about 1.5% short |
| Vertical | about 2.9% short |
The reason this survives so long in production is the first row. Operators check the system against a known horizontal feature, it agrees, and confidence is established. The error only appears on diagonals and verticals — and it appears as a small consistent bias, exactly the signature people attribute to "operator variation".
On a 40 mm vertical feature, 2.9% is 1.16 mm. That is not a subtlety. That is a scrapped batch.
Detecting it in ten minutes
- Image a circular artefact. A precision ring gauge, a reamed bore, a ball bearing — anything you trust to be round. Place it near the centre of the field.
- Measure it horizontally and vertically in pixels. Use the same method for both.
- Divide. The ratio of the two pixel spans is your aspect ratio. A truly round feature reading 500 px across and 515 px down tells you sy is 500/515 of sx — a 3% anisotropy.
- Confirm with a two-axis artefact. A cross graticule or a grid target lets you calibrate each axis directly against a traceable length rather than inferring one from the other.
- Check the part is actually perpendicular before you accept the result. Rotate the artefact 90° and measure again. If the anisotropy follows the artefact rather than staying with the axes, you have a tilt or a non-round artefact, not a pixel aspect problem.
Step 5 is the one people skip. Anisotropic calibration is the right fix for optics and sensors. It is the wrong fix for a part that is not sitting flat, because it will make the numbers agree while leaving the geometry wrong — and the next part, fixtured slightly differently, will disagree again.
How Measuret handles it
Anisotropic calibration is not a special mode in Measuret. It is the normal case. Every row in the calibration table has a Scale X and a Scale Y as separate columns, and the two calibration entry panels — "Calibration along X-axis" and "Calibration along Y-axis" — are presented side by side, so calibrating only one axis is a visible choice rather than an oversight.
Underneath, the calibration is a 2×2 transform rather than a single multiplier, so distances, angles, radii and every relationship between drawn shapes all inherit the correct treatment. There is no separate "anisotropic distance" tool, because there does not need to be one.
Two consequences are worth stating plainly:
- A round feature genuinely is elliptical in pixels under anisotropic calibration. This is correct, not a defect.
- Because angles are computed in calibrated space rather than pixel space, an angle measured on an anisotropic image is the real angle on the part — not the angle you would get by reading the pixels directly.
Anisotropic scaling handles axes that differ in magnitude. It does not correct lens distortion, perspective, or a part that is not perpendicular to the optical axis — those are different problems with different fixes.