Micron-level image measurement accuracy: why it drifts and the fix

When hole positions on machined parts must hold ±2 µm and the readings wander by 3 µm repeat-to-repeat, the first suspects are the machining, then the instrument. At this level both deserve suspicion: the measurement chain — temperature, edge quality, calibration, circle fitting — carries more error than the camera's pixels ever will. This guide is for process and quality engineers running camera-based gauging on machined features. You will put a number on thermal drift for your span, run three tests that split the remaining error, and finish with a fix order and an optics shortlist that holds single-micron repeatability.

Start from the numbers you already have

The requirement is ±2 µm on hole position. The gauge-maker's rule says the measuring instrument should carry a tenth of that uncertainty — about 0.2 µm — before a pass or fail means anything. Repeated runs on the same part show a 3 µm spread and a worst shift of 7 µm across the pattern.

That makes the gauge roughly 15 times too coarse to judge the process: 3 µm ÷ 0.2 µm = 15. And a part that measures short every time is showing a bias, not noise. Averaging more images cannot remove a bias; only geometry fixes do.

The sections below work through these in order, with the arithmetic that decides each one. The general method — measuring object size from a camera image — starts the same way: quantify the instrument before judging the part.

The number that sets the environment: thermal expansion

One term dominates micron-level dimensions on metal:

Steel grows α ≈ 11 µm per metre per °C — the standard tabulated value for carbon steel. Aluminium is about 23 µm/(m·°C), roughly double. The term is linear in all three inputs: double the span or the temperature change and the drift doubles.

This is why high-precision machining centers quote their accuracy for a ±1 °C room: on the span in the table's second row, ±1 °C is ±0.55 µm, a quarter of your tolerance. A shop that swings 10 °C between a cold morning start and mid-afternoon has spent 5.5 µm before the camera is switched on.

Worked example: a steel hole pattern drifting 3 µm through a shift

Inputs

• Measured span L = 50 mm — the hole-to-hole distance on the part.
• Tolerance ±2 µm on hole position — the inspection requirement.
• Steel α ≈ 11 µm/(m·°C) — standard linear-expansion tables for carbon steel.
• Gauge behavior: 3 µm repeat-to-repeat spread, 7 µm worst position shift — from the repeated runs.
• Reference: gauge blocks certified at 20 °C.

The rule

Every °C between the part and the 20 °C reference lengthens a steel span by 11 µm per metre of span. Bias appears when part, stage and blocks disagree in temperature.

The arithmetic

Growth per °C: 11 µm/(m·°C) × 0.050 m = 0.55 µm/°C.

Machine warm after cutting, ΔT = 3 °C: 0.55 × 3 = 1.65 µm.

Cold start to afternoon shift, ΔT = 10 °C: 0.55 × 10 = 5.5 µm.

Temperature that consumes the ±2 µm limit alone: 2 µm ÷ 0.55 µm/°C = 3.6 °C.

The result and the decision

A 10 °C swing moves a geometrically perfect span of this length by 5.5 µm — 38% more than the entire 4 µm tolerance band, and the same order as the observed 3-7 µm error. Treat ±3.6 °C as the hard ceiling: beyond it, a perfect part fails on temperature alone. Hold part, stage and blocks within 20 ±1 °C (±0.55 µm), soak parts before measuring, or log temperature and subtract α·L·ΔT from every span.

Test 1: repeatability with nothing else changing

Image the same part 30 times over 15 minutes without unclamping. The max-minus-min of each hole position is the instrument alone.

Accept the gauge when that spread stays under 0.4 µm — 10% of the 4 µm band — with a 0.2 µm uncertainty target. At 3 µm, stop tuning the process: the measurement system is the finding.

Then repeat morning and afternoon with the block, logging stage temperature. Predicted drift is 0.55 µm/°C × logged ΔT on steel; a measured shift matching the prediction convicts temperature. What averaging can do is cut the random remainder: averaging measurement values per part across repeated runs buys √N — 25 repeats cut random scatter 5× — but never a bias.

Test 2: the edges — burrs, chatter and chips

A burr ring blocks backlight at the rim and shrinks the fitted circle. A one-sided burr, the signature of tool deflection, drags the fitted center sideways — position error, not size error. Chatter waves the edge so the gray-level ramp changes run to run, and a threshold lands on a slightly different point each time. Chips sitting in the hole shadow the rim entirely.

Test it: export the gray-level profile across one edge, overlay 20 runs, and watch where the ramp moves. Deburr, ultrasonic-clean, dry, and run the 30-image repeat. If the scatter collapses, the edges were the cause. Edge position accuracy from image sampling shows how much of the budget the ramp width alone consumes.

Test 3: calibration temperature and the circle fit

Calibrate scale with gauge blocks at the temperature you measure at, and use precision-grade blocks: in this size class they carry ±0.2 µm, the right class for this job. The arithmetic bites in both directions — the calibration block sitting 4 °C warm is 0.55 × 4 = 2.2 µm longer than its certificate. Calibrate scale on it and a 20 °C part reads about 2.2 µm short on the same span, which is exactly the everything-reads-short symptom.

Then the fit. A threshold blob center moves when the threshold moves, and every center inherits the half-pixel offset in blob centroids unless the pixel origin convention is handled. The stable choice is the grayscale centroid of a dark circle or a sub-pixel fit to the edge ramp: published studies of machined holes report error reductions around 75% from sub-pixel refinement over plain thresholding. Sweep the threshold ±10% and watch the reported positions — a stable fit moves a fraction of a pixel, a fragile one walks.

Focus belongs in the same test: refocusing changes the ramp width, so set a repeatable focus position during calibration and leave it there.

Optics: remove the perspective term

An ordinary lens changes magnification with part distance; the relative error is roughly Δz/WD for working distance WD. A 0.1 mm height difference at 100 mm working distance is 0.1% — 50 µm across the pattern, 25 times your whole tolerance. A telecentric lens holds magnification constant across its depth of field, removing that term. When you need a telecentric lens for gauging covers the crossover; telecentric depth of field sets the height range you must stay inside.

Size the magnification to the span

Field first: 50 mm × 1.3 = 65 mm minimum FOV. Magnification follows from the sensor: 12.8 mm (a 1" sensor's width) ÷ 65 mm ≈ 0.20x. A 1" sensor at 0.166x covers 12.8 ÷ 0.166 ≈ 77 mm — about 13 mm of margin per side.

The LS-TCL-1422 is a bi-telecentric 0.166x lens with 258 mm working distance for 1" sensors on a C-mount.

Object-space pixel size is sensor pitch ÷ magnification. An illustrative 5 MP sensor with 2,448 pixels across 12.8 mm has a 5.2 µm pitch; 5.2 ÷ 0.166 ≈ 31 µm on the part. The pixel grid is not your limiter: a clean circle spreads its edge information over hundreds of boundary points, which is why center positions repeat well below one pixel. Object-space pixel size from magnification converts this for any camera.

77 mm ÷ 31 µm ≈ 2,480 pixels across — 5 MP class, the illustrative sensor above.

Single holes on their own want a tighter field: 1x on a 2/3" sensor covers about 8.8 mm. Three telecentric options cover the common cases:

All 3 share: Mount C-mount. Values as printed in each product's data sheet; open a part for its drawing and full specifications.

Check working distance against your fixture before ordering; the telecentric lenses range and the lens finder cover the alternatives.

Light and camera: buy edge stability, not resolution

Backlight the holes. A collimated light behind the plate with a telecentric lens in front makes the rim a sharp, threshold-stable edge; front diffuse light on a machined face drags the threshold around with surface finish. Collimated or diffuse backlight for edge gauging compares the two, and the collimated lights range is built to pair with telecentric optics.

The LS-CLM-0063 is a Ø100 mm telecentric collimated light, 5 V, with a 3-pin flat connector.

A 100 mm beam covers the 77 mm field with margin. Drive it from a light controller in constant-current mode and warm it up a few minutes before calibrating so the LED output is settled.

The camera matters least, but three choices still count: mono over color (no Bayer interpolation across your edges), global shutter (vibration immunity), and a sensor that fills the lens image circle — a 1" telecentric wants a 1" sensor. The LS-SMC-0017 is an 8.8 MP smart camera with a 1" sensor, C-mount, global shutter and an on-board X86 platform, in color and mono versions.

If you need more pixels or another interface, the smart cameras category and the camera finder cover the rest.

When the usual advice is wrong

More megapixels is the reflex, but pixel size is not the bottleneck: 31 µm object pixels reach a 0.2 µm budget through sub-pixel fitting and averaging, not sensor resolution, and a higher-resolution camera on a burr-rimmed hole still drifts.

Recalibrating more often can make things worse — each calibration bakes that session's expansion into the scale if the stage temperature moved between sessions.

Measuring on the machine with a touch-trigger probe is a legitimate alternative for hole centers — the part never leaves its thermal state — but it inherits the same 10:1 duty, so check it against gauge blocks too. When two instruments disagree beyond their combined budgets, send the artifact to an accredited lab for arbitration instead of tuning the process to the average.

Checklist and the fix order

Before you order:

• Telecentric, object-side or bi- — removes the perspective term that reads long or short with part height.
• FOV ≥ 1.3 × span — 65 mm minimum for this pattern; 77 mm at 0.166x is comfortable.
• Image circle ≥ sensor format — a 1" lens pairs with a 1" sensor.
• Collimated backlight, beam ≥ FOV — sharp, threshold-stable hole rims.
• Constant-current light drive — LED output settled before calibration.
• Mono, global-shutter camera — no Bayer interpolation, vibration-immune.
• Precision-grade gauge blocks in the right class — the reference must sit far below the 0.2 µm budget.
• Part soak at 20 °C plus ultrasonic cleaning — removes the two largest error sources before the camera sees the part.

The fix order in production:

1. Hold part, stage and blocks at 20 ±1 °C; soak parts after cleaning.
2. Deburr consistently, ultrasonic-clean, dry.
3. Calibrate scale with precision blocks at measurement temperature.
4. Fit circle centers with the grayscale centroid or a sub-pixel edge fit.
5. Verify on a gauge block: spread ≤ 0.4 µm before judging any part.
6. Log stage temperature with every measurement so a drift arrives with a number attached.

The decision, per case: scatter that follows the clock is temperature; scatter that cleaning collapses is edge quality; short readings by a fixed fraction are calibration scale; error that tracks part height is perspective; scatter that tracks the threshold is the fit. Work the chain from light to lens with the lens configurator and the lighting finder, read the applications guides for the surrounding workflow, and bring us your drawing through the contact page — every part carries a 3-year warranty minimum and expert support with the order.