Are you checking dimensions with calipers on a sample basis while your customer now wants 100% inspection right on the line? Camera-based dimensional measurement (machine vision metrology) is the usual answer, and the QC engineer's first question is always the same: how many microns of error can a camera really achieve, and can it replace calipers or a CMM? The honest answer is that it depends on the application. This article explains what drives accuracy so you can estimate feasibility before running a trial.
How does camera-based dimensional measurement work?
The camera captures an image of the part, software locates the part's edges with sub-pixel interpolation, and a calibration factor converts the distance between edges from pixels to millimetres. This is non-contact dimensional measurement: nothing touches the part, so there are no scratches, no deformation of soft parts and no probe wear. A single image can return several dimensions at once (diameters, hole spacing, lengths, angles), which makes it a natural fit for 100% inline dimensional inspection.
What determines the accuracy of machine vision measurement?
The accuracy figure on a camera datasheet does not tell you the real measurement result. System accuracy is a chain, and the weakest link sets the outcome.
1. Resolution: µm per pixel
Each pixel corresponds to a length on the part. The formula is µm/pixel = field of view (FOV) ÷ number of pixels along that axis. For a 5 MP camera with 2,448 pixels across (the format of common 5 MP area-scan sensors):
- 60 mm FOV: 60,000 µm ÷ 2,448 ≈ 24.5 µm/pixel.
- 30 mm FOV: ≈ 12.3 µm/pixel.
- 15 mm FOV: ≈ 6.1 µm/pixel.
A smaller FOV gives finer resolution, but the part must still fit in the frame. Sub-pixel edge detection can read an edge position to a fraction of a pixel, but the real gain depends on contrast, noise and image stability, so do not treat it as guaranteed when designing the system.
2. Telecentric lens
A standard lens shows near objects larger than far ones (perspective error) and adds distortion. For precise measurement, or for parts whose height varies, you need a telecentric lens: light enters parallel to the optical axis, so image size barely changes when the part sits slightly higher or lower. For example, per the manufacturer's data the OPT BP20-65-LQVTC telecentric lens has 2.0X magnification, a Ø9.20 mm FOV, a 65 ± 2 mm working distance, distortion below 0.1% and 0.10° telecentricity. Lenses like this suit small parts; larger parts need lenses with a bigger FOV, which are bulkier and more expensive.
3. Backlight illumination
To measure, the part's edges must be sharp and stable. A backlight (placed behind the part, with the camera looking back toward it) produces a crisp dark silhouette on a bright background, giving the highest edge contrast. When you only need the outline, a diameter, a hole or a slot, backlight is practically the default. If you must measure surface features (steps, shallow grooves), you need other lighting and accuracy is usually lower. See our machine vision lighting guide for how to choose.
4. Calibration
Calibration converts pixels to millimetres, usually with a calibration plate or a master part measured on a traceable instrument. An error here shifts every measurement in the same direction. Check it periodically, because vibration, temperature and camera removal can change the factor.
5. Part tolerance
A common metrology rule of thumb says the measurement method should be several times better than the tolerance, often about one tenth of the tolerance band. For a ±0.05 mm tolerance (a 0.1 mm band), aim for a measurement uncertainty around 10 µm. A 5 MP camera with a 15 mm FOV (≈ 6.1 µm/pixel) can get there with a well-chosen lens, lighting and calibration; a 60 mm FOV (≈ 24.5 µm/pixel) is unlikely to. The surest way to know is to measure real samples and assess repeatability and reproducibility.
Camera vs calipers, CMM and laser sensors
| Criterion | Calipers / micrometers | CMM | Laser sensor | Vision camera |
|---|
| Contact | Yes | Yes (probe) or optical | No | No |
| Speed | Slow, operator-dependent | Slow, batch-based, usually in a lab | Very fast, usually one dimension or one profile | Fast, many dimensions per image |
| 100% inline inspection | No | No | Yes | Yes |
| Complex 3D geometry | Limited | Excellent | Good for height/profile | Mostly 2D; 3D camera needed for height |
| Typical role | Spot checks, cross-checks | Reference standard, first-article checks | One fixed parameter | Multiple 2D dimensions on every part |
The key point: a camera rarely replaces a CMM completely. The CMM usually stays as the reference for calibration and cross-checking, while the camera handles the 100% inspection that a CMM cannot keep up with at line speed.
When should you use camera measurement? Three situations
- You need 100% inspection at line speed. Sampling with calipers misses isolated defects; a camera checks every part and logs data for traceability.
- Parts are soft, easily scratched or tiny. Gaskets, films, electronic components, thin plastic parts: non-contact measurement avoids damage.
- Many 2D dimensions on one part. Diameters, hole-centre distances, lengths and angles: one image replaces several manual measurements and removes operator-to-operator variation.
When should you not use it yet? Two cases
- The tolerance is too tight for the resolution you can realistically reach. When the tolerance band is only a few microns on a large part, a single camera is unlikely to cope; a CMM or a dedicated metrology system is the better choice.
- The feature is hidden from the camera or the geometry is complex in 3D. Deep bores, hidden faces and multi-axis 3D profiles suit a CMM or a 3D measurement system better. For more, see our page on 2D/3D camera vision metrology.
A quick calculation before a trial
A stamped metal part is 25 mm long with a ±0.05 mm tolerance. A 30 mm FOV keeps it fully in frame. A 5 MP camera gives ≈ 12.3 µm/pixel, which does not comfortably meet a 10 µm target. You have two options: a higher-resolution camera, or two images with a smaller FOV each. With a 20 MP camera at the same 30 mm FOV, the pixel count across the frame rises, so µm/pixel falls noticeably. Choose resolution and lens together; see how to choose an industrial camera for machine vision.
Checklist before you start
- The tolerance of each dimension and the largest part size.
- The minimum FOV, the resulting µm/pixel, and the lens type (telecentric or standard).
- Lighting: backlight for outlines, something else for surface features.
- Part presentation: vibration, height variation, conveyor speed.
- A calibration method and a master part measured on a reference instrument.
- How you will judge results: repeated measurements of the same sample, and comparison with the CMM on an OK/NG sample set.
Frequently asked questions
How many microns of error can a camera achieve?
There is no single figure. Resolution on the part (µm/pixel), lens quality, lighting, calibration and line conditions together decide it. The reliable way to find out is to test on real samples.
Can a camera fully replace a CMM?
Usually not. Cameras excel at 100% inline checks of 2D dimensions; the CMM remains the reference standard and handles complex 3D geometry.
Is a telecentric lens mandatory?
Not for every application, but use one when you need accurate measurement on parts with varying height or must reduce perspective distortion.
How BeeVision helps
BeeVision is a Cognex partner in Vietnam and also supplies Basler cameras and OPT and Smart Vision Lights lenses and lighting for measurement tasks. Our engineers do not promise accuracy before seeing a sample: send your product samples and the required tolerances, and BeeVision will assess feasibility free of charge and run a measurement trial on your real samples, returning a recommended setup together with actual repeatability data.
Follow BeeVision on LinkedIn and YouTube, and send samples through the sample POC page to find out whether your measurement task can be done with a camera.