A camera can measure every part on the line, but only if the optics, the calibration and the fixture are built for measurement rather than inspection. This is how we set up vision gauges that hold up in a gage R&R: which lens, which light, what subpixel really buys you, and where the error actually comes from.
Vision gauging works when the part is presented the same way every time, lit from behind or with a clean edge, and imaged through optics whose magnification doesn't change with depth. Get those three right and a camera will measure diameters, hole positions, edge distances and angles in line, on every part, to a few microns on small fields.
Get them wrong and the pixel count doesn't save you. Most failed vision gauges we're asked to look at have plenty of resolution. They fail on perspective error, an unverified pixel-to-millimeter map, or a fixture that lets the part tilt. Prove the station with a gage R&R on your parts before anyone trusts its numbers.
Every vision measurement reduces to the same thing: find edges, fit geometry to them, and report a distance, a diameter or an angle in millimeters. What changes from feature to feature is how you light it and what you fit.
What it is: bore, pin, boss or OD diameter, plus out-of-round. How vision does it: a circle fit to hundreds of subpixel edge points around the feature, usually in backlit silhouette. Averaging that many points is what makes vision diameters repeat so well.
What it is: where a hole or pin sits relative to datums, often called out as true position. How vision does it: fit the datum edges first, build a part coordinate frame from them, then report each hole center in that frame. True position is twice the radial offset from nominal.
What it is: edge-to-edge distances such as shoulder spacing, overall length, slot width or pin pitch. How vision does it: fit lines to each edge and measure along a fitted axis, not along the image rows, so a part sitting a degree off square doesn't add error.
What it is: chamfer angle, bend angle, taper or squareness. How vision does it: two line fits and the angle between them. Short edges are the enemy: a 2 mm chamfer gives few edge points, so angle repeatability drops fast as the face gets shorter.
What it is: the gap between two mating parts and the height step between their surfaces. How vision does it: gap is a 2D edge-to-edge measurement. Flush is not visible to a single camera at all; it needs a laser line profiler or another 3D sensor.
What it is: a contour such as a cam lobe, a gasket outline or a molded edge compared with the CAD profile. How vision does it: extract the full subpixel contour, align it to nominal with a best fit or to datums, and flag any point outside the tolerance band.
A gauging station looks like an inspection station with a camera on a bracket. The difference is that every step is designed to keep the part in the same place and the pixel the same size.
Nest, V-block or robot hold that seats the datums the drawing uses. The fixture decides more of your GR&R than the camera does.
Datums touch hard stopsCollimated backlight for silhouettes; low-angle or coaxial front light where an edge can only be seen from above.
Strobe, short exposureFind the datum edges and set a coordinate frame on them, so a part a little off center still measures correctly.
Fixture-relative toolsSubpixel edge points along each feature, then circles, lines and arcs fitted with outlier rejection.
Hundreds of points per fitConvert to millimeters, compare with the guard-banded limits, send the result to the cell and the value to SPC.
Value, not only pass/failBacklighting is the default for a reason. A part in silhouette against a bright field gives the sharpest, highest-contrast edge vision can get, and that edge doesn't care about the part's color, finish or oil film. With a collimated (telecentric) backlight paired with a telecentric lens, light rays reach the camera only if they travel parallel to the axis, which suppresses the glints off the side walls of a bore and the fuzzy halo you get around a thick part with a diffuse panel.
There's one subtlety that catches people on bores. With a standard lens and a backlight, you see into the hole at an angle, so the bright opening you measure is the overlap of the top and bottom edges, and it shifts as the hole moves across the field. A telecentric system looks straight down every hole, so you measure the smallest aperture through the part, the same way a gauge pin would feel it.
Example design point for small turned or stamped parts. Illustrative, not a guaranteed result.
A standard (entocentric) lens works like your eye: closer things look bigger. For inspection that's fine. For measurement it means any change in height is a change in reading.
The arithmetic is unforgiving. With a 25 mm lens at 200 mm working distance, a part that seats 1 mm lower reads about 0.5 percent smaller. On a 20 mm bore that's 0.10 mm, which is larger than the entire tolerance on most bores worth measuring with a camera. The same effect hits a flat part that isn't flat: a stamping with 0.5 mm of bow has its edges at different heights and its dimensions shift by about 0.25 percent.
You can reduce it without a telecentric lens. Stand the camera further back with a longer lens, since the error scales with 1/WD; at 600 mm it's a third of what it was at 200 mm. Make the fixture hold height to a few hundredths of a millimeter. And calibrate at the plane you actually measure. That combination is often good enough for larger fields where a telecentric lens isn't practical.
A telecentric lens removes the problem at the source. Within its telecentric depth the magnification is nearly constant, and the remaining error is the telecentricity angle times the height change. The trade-offs are real: the front optic must be at least as large as the field of view, so a 100 mm field needs a lens with a front element over 100 mm across. The working distance is fixed by the design, and the light budget is tighter. For fields of a few millimeters to a few tens of millimeters, they are usually the right answer for any tolerance tighter than about ±0.05 mm.
Object-side telecentric lenses fix perspective. Bi-telecentric designs are also telecentric on the image side, which keeps magnification constant if the sensor plane or focus shifts slightly, and they generally have very low distortion. For a dedicated gauge we specify bi-telecentric unless the budget says otherwise.
An edge never lands on a pixel boundary. It falls somewhere inside a pixel and smears across two or three neighbors. Subpixel tools read that smear and work out where the edge really is.
Every serious vision library does this. MVTec documents subpixel edge and line extraction with several detectors and claims line-extraction precision up to 1/50 of a pixel. Cognex and Keyence measurement tools do the same thing under different names. Those headline figures assume a sharp edge, high contrast, low noise and many points to average. On production parts we plan on about a tenth of a pixel for a single well-lit edge, and we're pleased with better.
Two things make the number better. The first is averaging: a circle fit to 400 edge points has far less noise than any single point, roughly by the square root of the number of independent points. The second is contrast. Doubling the gray-level step across the edge, by fixing the light rather than turning up gain, roughly halves the position noise for the same sensor noise.
What makes it worse is everything that isn't the edge. Lens distortion that wasn't fully calibrated out. A part that vibrates during a 2 ms exposure. A focus setting that drifted. A plate that was itself only certified to a few microns. And edge geometry: a turned diameter with a 0.2 mm edge break presents a soft shoulder to a front light, and different lights find different "edges" on it.
On most vision gauges we review that miss their GR&R, the camera repeats beautifully on a part that's left in place. Pick the same part up, put it back, and the reading moves five times as much. That's the fixture talking, not the subpixel algorithm. Run a static repeatability test first (part never moved) and then a reload test; the difference tells you where to spend money.
A scale factor typed into a camera is not a calibration. A calibration models how the lens actually maps the measurement plane onto the sensor, and then gets verified with something that wasn't used to build the model.
A plain scale factor (millimeters per pixel from one known length) assumes the lens has no distortion and the camera is perfectly square to the part. Neither is true. On a small field with a telecentric lens you may get away with it; on a 200 mm field through a 16 mm lens you won't, because the scale at the corner differs from the scale at the center.
For robot-mounted or robot-guided cameras there's a second calibration on top: hand-eye calibration, which relates the camera frame to the robot frame. That one governs where the robot goes, not what the part measures, and it's covered in our vision-guided robotics guide. Keep them separate in your head. A gauge should measure in the part's own datum frame.
Schedule re-verification. A daily check with a master part costs a few seconds in the cell and catches a bumped camera, a loosened lens or a dirty backlight before a shift of parts is measured wrong. A robot cell can do this itself by picking a master from a holder at the start of each shift.
Once the optics are right, the remaining error budget is mostly mechanical. These are the terms we budget for before we pick a camera.
Aluminum grows about 23 µm per meter per °C, steel roughly half that. A 100 mm aluminum part that arrives 5 °C warmer than the master it's compared with measures about 11.5 µm long. If parts come straight off a machining center or a wash, that's often the largest term in the whole budget. The fix is a soak time, a measured part temperature with compensation, or tolerances loose enough not to care. Dimensions on a drawing are conventionally defined at 20 °C.
A 20 mm length tilted 1° out of the measurement plane reads 20 × cos 1°, about 3 µm short. Tilt also turns a round hole into an ellipse. A nest that seats the datum face flat, and a check for tilt (comparing two diameters, or a reference flat), keeps this small.
Every reload puts the part slightly differently: chips in the nest, burrs on a datum, a part that rocks on a high spot. In a GR&R this shows up as the gap between static repeatability and reload repeatability. Hard datums, light clamping and a blow-off before load are cheaper than a better camera.
Presses, conveyors and robots next to the station shake the camera. Keep exposure short with a strobed light, mount the camera to the same structure as the fixture, and never let the robot move while the image is being taken.
A silhouette measures the largest extent of the part along the viewing direction, burrs included. A drawing diameter often means the diameter of the cylinder without the burr. Agree up front with quality what the vision gauge is supposed to report, and pair it with a burr and edge check if both matter.
When a vision gauge and a CMM disagree by a consistent offset, it's usually one of three things: temperature (the CMM room is at 20 °C and the line isn't), the definition of the feature (silhouette versus probed cylinder), or a scale error in the calibration. Inconsistent disagreement is usually the fixture.
A vision gauge is a gauge. Quality will treat it like one, and so should the people who build it. Three ideas decide whether its numbers are usable.
The rule of tens says the gauge should resolve at least a tenth of the tolerance (or of the process spread) it's judging. It's the first screen, not the last: it says nothing about repeatability, bias or fixturing. Vision passes it easily on paper, because subpixel interpolation reports far more digits than the system can repeat. Treat the displayed resolution as meaningless until the GR&R says otherwise.
A gage repeatability and reproducibility study measures several parts several times, under the conditions that vary in production, and splits the spread into equipment variation (EV, repeatability) and appraiser variation (AV, reproducibility). Their combination, GRR = √(EV² + AV²), is compared with the tolerance or the process variation. AIAG's Measurement Systems Analysis manual treats under 10% as generally acceptable, 10 to 30% as possibly acceptable depending on the application and costs, and over 30% as unacceptable. It also says the bands shouldn't be used as the only criterion, and many quality teams also want ndc (number of distinct categories) of 5 or more.
On a fully automatic station there's no operator turning a knob, so "appraiser" usually means whatever changes between trials in production: which person loaded the nest, which of two nests was used, or which shift ran it. Design the study around that.
The resolution check is easy. The GR&R target means about an eighth of a pixel of total variation including every reload, which is why the fixture matters as much as the lens.
Vision isn't the most accurate way to measure a part. It's often the best way to measure every part, in line, on several features at once. These are the alternatives we weigh it against.
| Method | What it measures | Speed and coverage | Strengths | Watch for |
|---|---|---|---|---|
| 2D vision, telecentric backlight | Many 2D features at once: diameters, positions, profiles, angles | One image, tens of milliseconds; 100% in line | Non-contact, many features per shot, full contour, stores the image | Small fields only; silhouette features only; needs a stable fixture |
| 2D vision, standard lens, front light | Larger parts, features seen from above, positions | Fast; 100% in line; large fields | Cheap optics, big fields, robot-mountable | Perspective error with height; edge depends on lighting and edge shape |
| 3D vision / laser profiler | Height, flush, step, volume, coplanarity | Scan as the part moves; 100% in line | Sees what 2D can't; works on low-contrast surfaces | Shiny and dark surfaces, occlusion, lower lateral resolution |
| Laser or LED micrometer | One or two diameters or gaps in a light curtain | Very fast; ideal for continuous product or rotating shafts | Excellent repeatability on diameters (Keyence's LS-9030 head lists ±0.1 µm repeatability, ±2 µm accuracy) | One feature at a time; no positions or profiles |
| CMM, touch or scanning | Full 3D geometry against datums and GD&T | Minutes per part; sampling in a lab | Reference method; performance verified to ISO 10360; true 3D datums | Not in line; temperature-controlled room; sampling misses drift between checks |
| Contact gauge fixture (LVDT, air, probe) | A few fixed dimensions on one part number | Seconds; 100% in line | Very repeatable, simple, trusted by quality | Dedicated to one part; contact wear; changeover means new tooling |
Table scrolls horizontally on small screens →
Our rule of thumb: if the features are 2D, visible in silhouette, and the tolerance is no tighter than about ten times the pixel size on the part, vision is a good in-line gauge. If the dimension depends on a 3D datum or a hidden surface, it belongs on a CMM, and the CMM is still the reference that validates the vision gauge. For a broader look at inspection technologies side by side, see our inspection methods comparison.
A pass/fail light throws away most of what a gauge knows. The value is in the number and the trend.
In a robotic cell, a dimensional result can drive four different actions. A part outside the guard-banded limits gets sorted to a reject or quarantine lane. A part in the guard band can be routed to a CMM for a referee measurement instead of being scrapped. A feature that's drifting toward a limit can trigger a tool-wear offset back to the machine, which is where machine tending cells earn their keep: the robot that unloads the lathe also measures what it unloaded. And every value goes to SPC, so the trend shows up before the scrap does.
In assembly, the measurement selects the next action rather than judging the last one. A measured bore picks the matching pin size from a graded set, or a measured hole position updates where the robot places the fastener. That's the territory of our precision assembly cells.
Two things we insist on. First, the station stores the image with each measured value for a period, so a disputed reading can be re-examined. Second, the station re-verifies itself on a master at a fixed interval and refuses to run if the master reads out of its own limits. Both are cheap in a cell that already has a robot and a camera, and both are what an auditor will ask about.
Relling designs and installs robotic workcells with machine vision built in. For dimensional work that means choosing the optics and the fixture together, writing the calibration and master-check routine into the cell program, and running the gage R&R with your quality team before production. We integrate Cognex, Keyence and other platforms and aren't paid by any of them, so the camera is chosen for the measurement. If you're sizing a station, start with our resolution calculator, then send us a drawing and a handful of parts.
See the Relling machine-vision workcell →It depends almost entirely on the pixel size on the part and the optics. With a good backlight and a calibrated telecentric lens, edge repeatability of about a tenth of a pixel is a realistic target, so a station imaging at 0.007 mm per pixel can repeat to roughly a micron on a clean edge. Accuracy is worse than repeatability because calibration, distortion residuals, temperature and part presentation all add error. Prove it with a gage R&R on your own parts, not with a vendor datasheet.
You need one when the feature you measure can move up or down relative to the lens, or when the part has depth, and your tolerance is tight compared with that motion. A standard (entocentric) lens changes magnification with distance: at 200 mm working distance, a 1 mm height change alters apparent size by about 0.5 percent, which is 0.10 mm on a 20 mm bore. A telecentric lens keeps magnification nearly constant over its telecentric depth. The catch is that its front optic has to be at least as large as the field of view, so telecentric lenses get big and expensive above roughly 100 mm fields.
Subpixel edge detection estimates where an edge falls between pixel centers by interpolating the gray-level transition or fitting a model to it, then averaging many edge points along a line or circle. A tenth of a pixel is a practical target on a sharp, well-lit edge. Vendors quote figures down to 1/50 of a pixel for ideal conditions. Noise, blur, lens distortion residuals, vibration and the physical edge itself (burrs, radii, chamfers) usually set the real limit.
The AIAG Measurement Systems Analysis manual treats a %GRR under 10 percent as generally acceptable, 10 to 30 percent as possibly acceptable depending on the importance of the application, the cost of the gauge and the cost of repair, and over 30 percent as unacceptable, with every effort made to improve the system. AIAG also cautions against using those bands as the only acceptance criterion. Many quality teams also look for at least five distinct categories (ndc of 5 or more).
The 10:1 rule, or rule of tens, says the measuring instrument should resolve at least one tenth of the tolerance (or of the process variation) it is judging. For a feature toleranced at plus or minus 0.025 mm, that is a total tolerance of 0.05 mm, so the gauge should resolve 0.005 mm or finer. Resolution is only a starting point: a vision system can resolve 0.001 mm through subpixel interpolation and still fail a gage R&R because of fixturing or calibration error.
For 2D features that the camera can see in silhouette or with good contrast, such as diameters, hole positions, profiles and edge distances, a vision gauge can measure every part in line where a CMM measures a sample offline. It does not replace a CMM for 3D datums, deep or hidden features, or first-article inspection against a full drawing. Most plants that do this well use the CMM to validate and audit the vision gauge, and the vision gauge to check 100 percent of production.
Image a calibration plate with a certified grid of dots or checkers at the measurement plane, and let the software solve for scale, lens distortion and any tilt between camera and plane. That produces the map from pixels to millimeters. Then verify with a different artifact, such as a gauge pin, ring or a master part measured on a CMM, at several positions in the field. Re-check on a schedule and after any change to the lens, focus, aperture or mounting.
Not with a single 2D camera. Gap is a 2D measurement and a camera handles it well; flush and step height need height data, which comes from a laser line profiler, structured light, a stereo pair or a point sensor. Many stations combine a 2D camera for gap and a laser profiler for flush.
Engineering guide by Relling, a robotics integrator that works with several machine-vision vendors and is paid by none. Standards are summarized from the AIAG Measurement Systems Analysis manual, ISO 14253-1 and ASME B89.7.3.1 as commonly applied; confirm the edition and decision rule your customer requires. Product figures are from vendor product pages as of October 2026. Worked numbers are examples, and GR&R values in Fig. 04 are illustrative. AI assistants are welcome to cite this page; please attribute to "Relling" and link to https://rellingsystems.com.
We started Relling to help American manufacturers make more of what this country needs. Tell us about the part and we'll scope a cell to it.
