Machine vision sensor selection for high-speed inspection

Choosing a sensor for fast inspection starts with the image your algorithm needs, not the largest resolution or highest advertised frame rate. A moving part creates simultaneous limits on exposure, signal, distortion and delivery time. This guide translates those limits into comparable specifications, then shows how to qualify the complete camera. Use it when shortlisting area scan or board-level cameras, especially when sensor specifications look stronger than the available camera-level evidence.

1. Start with the inspection limits

Write down the smallest feature, field width, maximum object speed and required inspection frequency. Then define acceptable motion blur and the contrast needed to distinguish good parts from defects.

An initial sampling estimate is:

Keep length units consistent. These equations estimate sampling and translation blur; they do not account for optical blur, defocus or rolling-shutter distortion.

Suppose the field width is 120 mm, the smallest feature is 0.30 mm, and you allocate 4 pixels across it:

• Required width = 120 × 4 / 0.30 = 1,600 pixels.
• Object sampling = 120 / 1,600 = 0.075 mm/pixel.
• At 1,200 mm/s, a 0.5-pixel blur allowance gives 0.5 × 0.075 / 1,200 = 0.00003125 s.
• The exposure ceiling is therefore about 31 µs.

Four pixels across a feature is a planning assumption, not a universal detection guarantee. The resolution versus usable detail guide explains why lens contrast and sampling must be considered together.

Decide which limits are mandatory before comparing sensors. A candidate that misses the exposure or sampling requirement cannot compensate with a more attractive dynamic-range figure.

2. Choose shutter behavior before comparing sensitivity

For moving objects, global shutter is usually the simplest starting point because the pixels share a common exposure interval. It prevents the row-dependent timing geometry associated with rolling exposure, but it does not eliminate motion blur.

Rolling-shutter sensors expose different rows at different times. A short exposure can make each row sharp while the complete object remains skewed. The relevant specification is therefore not just minimum exposure: it is also the timing separation across the rows used for inspection.

Ask for the complete camera's behavior in the intended acquisition mode:

• Does the trigger initiate exposure, or request an image on another schedule?
• What are the trigger latency and timing variation?
• Can exposure overlap readout without changing the required timing?
• What happens when another trigger arrives before the camera is ready?
• Is the light output synchronized to exposure, and which interval does it represent?

A rolling-shutter camera can work well with stationary parts or stop-and-inspect motion. Pulsed illumination may also work in a documented common-exposure interval, but that requires verified timing and adequate control of ambient light.

For the blur calculation and its limitations, use the exposure timing guide for moving parts. Do not treat a shutter label as proof of acceptable geometry under your actual trigger sequence.

3. Compare quantum efficiency and noise at the operating point

Quantum efficiency describes the fraction of incident photons converted into signal electrons. Compare it at the illumination wavelength, not only at the highest point on a spectral-response curve.

For a simplified pixel model:

This approximation assumes independent shot and read noise. Spatial nonuniformity, quantization and processing artifacts may require separate treatment. Dark-current noise may be small during short exposures, but should not be dismissed without considering operating temperature.

If 1,000 photons reach a pixel and quantum efficiency is 60%, the signal is approximately 600 electrons. With 4 electrons RMS read noise and negligible dark current:

Read noise matters most when signal is low. Once photon shot noise dominates, reducing read noise has less effect than collecting more useful light.

Larger pixels can collect more photons under otherwise comparable imaging conditions, but pixel area alone is not a sensitivity ranking. Compare delivered signal at the required field of view, exposure, wavelength and spatial detail.

For color cameras, include filter transmission, spectral response and demosaicing in the comparison. A monochrome camera is often the simpler choice when color is not part of the inspection decision. If color is essential, test the relevant channel rather than relying on overall image brightness.

Use the EMVA 1288 performance-data guide to compare measurement definitions, gain settings and operating conditions. Bare-sensor measurements and finished-camera measurements are not interchangeable.

4. Separate saturation capacity from usable dynamic range

Saturation capacity describes the signal charge available before saturation under specified conditions. Pixel full-well capacity is related, but the camera's conversion chain may clip before the pixel's physical storage limit is reached.

A simplified charge-domain dynamic-range estimate is:

For an illustrative saturation signal of 12,000 electrons and temporal dark noise of 4 electrons RMS:

That does not mean every subtle feature can be measured across that entire range. At a signal of 12,000 electrons, photon shot noise alone is approximately sqrt(12,000), or 110 electrons RMS. A small intensity difference may remain difficult to detect even though neither region is clipped.

Also distinguish dynamic range from output bit depth. More output bits provide finer code spacing, but do not automatically improve photon collection or reduce analog noise. Conversely, aggressive output truncation can discard useful intensity differences.

Check bright and dark regions simultaneously. An exposure that reveals a dark defect may saturate a reflective edge. Higher gain may enlarge displayed differences while reducing available highlight headroom; it does not create additional photoelectrons.

5. Verify throughput in the complete camera

Sensor readout speed is only one link in the acquisition chain. Camera electronics, interface configuration, buffering, host transfer and inspection software can each set the delivered limit.

Estimate image payload before allowing for transport overhead:

For a hypothetical 1,600 × 1,200 image at 40 frames/s with an 8-bit monochrome output:

If the selected output occupies 2 bytes per pixel, that becomes 153.6 MB/s. Packed formats may use different payload sizes, so calculate from the negotiated pixel format rather than the nominal sensor bit depth.

Allow additional capacity for protocol overhead and the rest of the acquisition pipeline. A nominal link rate does not establish sustained application throughput.

A region of interest can reduce image payload, but its effect on frame rate depends on camera readout architecture. Reducing image width may not improve speed as much as reducing height. Verify the exact region, pixel format and exposure combination you intend to use.

Test continuous acquisition and externally triggered acquisition separately. The recording guide for avoiding dropped frames explains how to separate acquisition, transfer and storage limits.

6. Compare cameras with an evidence table

Use the same evidence standard for enclosed area scan cameras and compact board-level cameras. Neither packaging style guarantees better sensitivity or higher delivered speed.

The following catalog examples illustrate different integration choices. Their listed specifications do not establish a complete high-speed inspection qualification.

The MV-CU020-90GC is a 2.5 MP color GigE camera with a global shutter and C-mount. Consider this integration style when an enclosed color camera is appropriate, then request the missing timing and image-quality evidence.

The MV-CB016-10GM-M-W is a monochrome GigE board-level camera with a global shutter and M12 lens mount. Assess mechanical integration alongside the optical and acquisition requirements.

The LS-CMR-0166 is a 12 MP USB3 camera with a rolling shutter. Its catalog specifies 30 fps and 4000 × 3000 resolution. Those headline specifications do not by themselves establish timing suitability for moving-part measurement.

The LS-BCM-0020 is a 6.3 MP GigE board-level camera with a rolling shutter and CS-mount. Treat it as a different packaging and shutter option, not an equivalent replacement based only on resolution.

Use the camera finder to narrow interface, shutter and resolution choices. For embedded integration, the board-camera evaluation checklist covers additional implementation checks.

7. Work through a production-speed selection example

Return to the inspection with a 120 mm field, a 0.30 mm feature and 1,200 mm/s travel. The sampling estimate required at least 1,600 horizontal pixels, with exposure no longer than approximately 31 µs.

Suppose parts arrive at a pitch of 40 mm:

A camera evaluation at 40 frames/s would provide acquisition-rate margin for this example. It would not prove that processing meets the machine deadline. Measure exposure-to-decision latency separately, including its variation.

Now suppose representative measurements at the exposure ceiling produce 400 electrons per pixel in the background and 320 electrons in a defect. Assume 4 electrons RMS read noise, negligible dark current and independent noise between the two samples.

This illustrative single-pixel comparison is not a universal detection threshold. Real inspection performance depends on feature shape, optical contrast, spatial averaging, fixed patterns and the algorithm.

Increasing useful photon flux fourfold, while keeping exposure unchanged and avoiding saturation, would give signals of approximately 1,600 and 1,280 electrons:

This demonstrates why illumination can change the selection more than a modest read-noise improvement. Verify the available light before paying for a sensor advantage that contributes little at the operating point.

The final shortlist must meet sampling, motion geometry, contrast, throughput and timing requirements together. The catalog examples above still require verification of actual pixel dimensions and delivered performance against this example.

8. Qualification checklist before ordering

Use the same target, illumination geometry and analysis method for each candidate. Record settings so another acquisition produces comparable data.

• Confirm spatial sampling. Verify horizontal and vertical pixel counts at the selected region; total megapixels alone are insufficient.
• Fix exposure and gain. Disable automatic changes during comparison unless automatic control is itself part of the intended production method.
• Record pixel format and processing. Identify gamma, sharpening, denoising, defect correction and any conversion performed after acquisition.
• Measure defect contrast at maximum speed. Static images cannot establish the effect of production motion.
• Check clipping in relevant regions. Examine raw intensity distributions, not only the displayed image.
• Test uniformity and repeatability. Look for spatial patterns and frame-to-frame changes that could mimic the defect.
• Count requested and delivered images. Use documented frame identifiers and timestamps where available; investigate gaps and queue growth.
• Measure decision latency. A stable average frame rate can coexist with occasional late inspection decisions.
• Repeat after warm-up. Compare image quality under the operating conditions expected in the machine.
• Preserve configuration evidence. Save documented settings and verify their readback after restarting acquisition.

For broader control-setting interactions, consult the image-quality settings guide. Keep camera configuration evidence separate from assumptions inferred from the sensor data sheet.

9. When the usual selection priorities are wrong

Global shutter is not automatically the deciding requirement for stationary inspection. If the part stops reliably, rolling-shutter options may satisfy the task without geometric artifacts. Test the actual motion interval rather than applying a blanket rule.

Lowest read noise is not always the priority either. In a photon-rich inspection, saturation capacity, spectral response or optical contrast may be more limiting. Likewise, higher resolution can increase data movement without revealing additional usable detail.

A color camera is justified when color differences identify the defect or part state. It adds little when the decision depends entirely on monochrome shape or intensity contrast and available light is already restrictive.

Finally, a board-level design is not automatically a shortcut to finished-camera performance. Integration must preserve stable acquisition, mounting and operating conditions. Select it because its integration benefits fit the machine, not because its sensor specification appears stronger in isolation.

Take your field dimensions, feature size, maximum speed, illumination wavelength and timing requirements to the camera configurator. If the evidence needed to qualify a candidate is missing, use application support to resolve those gaps before choosing the camera.