Accept vision inspection cameras on measurements taken where they will work, not on the datasheet. That means five checks, each under one documented configuration: a slanted-edge MTF measurement at the real working distance and under the production lighting; signal-to-noise ratio and dynamic range measured the way EMVA 1288 defines them; a flicker test against the plant’s actual illumination; a check of sharpness against machine vibration; and the camera’s behaviour across the temperatures and humidity the site reaches. The datasheet tells you which cameras are worth testing. The measurements tell you which one to buy and whether the installation passes.
A datasheet number is a filter for which cameras to test, not evidence that the installed system works. The rest of this page covers the order of the tests, how each is set up, and the traps that make two honest measurements disagree. Three of them are processed versus RAW output, saturation capacity versus full-well capacity, and the quality of the chart.
The test sequence for an industrial camera for machine vision
The order matters because some effects hide others. Flicker can look like noise, and vibration can look like a soft lens. So we would establish the sensor’s own noise floor first, then optics, then the plant-driven effects one at a time.
| Step | What you measure | Method | The trap |
|---|---|---|---|
| 1. Read the datasheet | Claimed SNR, dynamic range, full well | Check every figure’s definition and operation point | Full-well capacity used where saturation capacity applies inflates dynamic range |
| 2. Sensor noise | SNR curve, dynamic range, dark current, DSNU/PRNU | EMVA 1288 Release 4 Linear photon transfer method | Invalid if the camera debayers, denoises or sharpens before output |
| 3. Optics on the rig | MTF50 / MTF50P per field position and channel | ISO 12233 slanted edge at 4:1 contrast | Sharpening makes slanted-edge results optimistic |
| 4. Illumination | Percent Flicker, frame-to-frame signal variation | PWM sweep on the bench, shift-long log on the line | Auto exposure left on corrupts the measurement |
| 5. Vibration | MTF and detection accuracy by machine-cycle phase | Static and moving targets compared across press phases | Test it separately from the flicker log |
| 6. Environment | Noise vs temperature; condensation risk | Temperature cycling with functional tests at the extremes | Passing an IP67 water-ingress test does not exclude condensation |
| 7. Decide | Gates, measured scores, cost, lead time | Gate first, then weigh the measurements | Stopping short of the site’s temperature extremes can miss degradation |
How do you check a camera datasheet’s SNR and dynamic-range claims against your own measurements?
You check them by re-measuring with the same method and definitions the datasheet used, with the camera in a linear output mode and every setting fixed, and then comparing like with like. For machine vision cameras that method is EMVA 1288, the European Machine Vision Association’s unified method to measure, compute and present specification data for cameras and image sensors. Datasheet figures are where the shortlist starts; the bench results are what the sourcing decision rests on.
LynxBenchAI measures AI throughput, latency, memory and energy across accelerators, so the hardware that runs the inspection model is compared on measured figures, just as the camera’s datasheet claims are; the first step is to benchmark the candidate accelerators.
The EMVA 1288 protocol in brief
EMVA 1288 Release 4.0 Linear supersedes Release 3.1 and applies only to cameras with a linear photo response. The camera must fit the standard’s linear model: if preprocessing such as debayering, denoising or edge sharpening breaks the white-noise assumption, Release 4 Linear cannot be used and the computed parameters are meaningless. Characterise it in its raw, unprocessed mode, or accept that the numbers you get are not EMVA numbers.
The setup illuminates the sensor without a lens from a homogeneous source at f-number 8, using narrow-band irradiation, an integrating sphere recommended and non-uniformity recommended at no more than 3%. The sweep uses at least 50 roughly equally spaced exposure levels (as few as 9 for production measurements), with two images per level plus two dark images at each exposure time. Plotting the noise variance against mean signal gives the system gain from the slope and the dark noise from the offset; this is the photon transfer method. From the same data the standard defines SNR as mean signal above dark divided by temporal noise, maximum SNR as the square root of the saturation capacity in electrons, and dynamic range as saturation divided by the absolute sensitivity threshold.
Dark current is measured from at least six equally spaced exposure times, and its temperature dependence from at least six equally spaced temperatures across the camera’s operating range. The irradiance calibration itself has a typical absolute accuracy of 3% to 5%, and the reference photodiode should be recalibrated at least every second year, which sets the minimum systematic error of anything expressed in photons. A careless bench biases every downstream number, so record the calibration state with the results.
A datasheet-reading checklist: why the same sensor gives different numbers
Definitions and measurement conditions move datasheet figures, so check them before suspecting a defective camera. Before calling a gap, check these:
- How full well is defined. Using the sensor’s full well instead of the usable, linear full well gives an artificially high dynamic range. Adimec’s advice is to look for a definition every time and ask the manufacturer when none is given.
- Saturation capacity versus full well. EMVA 1288 saturation capacity is normally lower than full-well capacity because the digital signal clips before the pixel physically saturates.
- Which dynamic-range definition. EMVA uses saturation over the absolute sensitivity threshold; ISO 15739 uses saturation luminance over the luminance where SNR equals 1, and allows that minimum to be extrapolated. Adimec defines it against readout noise in the dark.
- Same conditions. Dynamic-range values from different cameras are only comparable when measured under the same conditions.
- Operation points. An EMVA 1288 datasheet must state gain, offset, exposure and camera mode so that anyone can repeat the measurement; unstated parameters are assumed to be factory defaults.
- Which release. A real Basler EMVA report for the acA1920-48gm, for example, gives a maximum SNR of 38.9 dB (±0.13 dB) and was measured to release A1.03.
Adimec’s published figures show why the definition matters. Averaging 45 manufacturing tests on S-25 cameras, Adimec reports a full well of 16,119 e⁻, dark noise of 34 e⁻, dynamic range of 53.7 dB and SNR of 42.0 dB. Adimec also states that the linear part of the full well only extends to 13 ke⁻.
How big is the gap, and what to measure on an HDR sensor
We would distrust any typical datasheet-to-field gap quoted without the definitions behind it. The photon transfer workflow on dark and uniformly lit raw image sets yields SNR1, saturation level and dynamic range, but evaluating linearity in HDR mode is harder because the light source must keep a steady narrow-band spectrum across the full range. The figure to ask for is the SNR at each exposure transition, not only the total range. One 2021 Image Sensor Workshop paper on an automotive sensor reports, at room temperature, 140 dB dynamic range with every total-SNR transition at or above 30 dB.
Motion artefact belongs in the same assessment. Combining data from several time-shifted integrations produces motion artefacts, whereas a single long integration gives motion blur but no such combination artefacts. Motion blur grows with how far parts move during the exposure. That makes HDR motion artefact a scene-dependent measurement to repeat on your own line, not a bench number to carry over.
Which chart gives a defensible MTF measurement at the real working distance?
For a defensible number, use an ISO 12233 slanted-edge target with 4:1 edge contrast. Image it at the real working distance under the production ring lights, and report MTF50 or MTF50P per field position. Keep a bar target such as USAF 1951 for a quick visual check, because a limiting-resolution reading correlates less well with perceived sharpness than MTF50 does.
LynxBenchAI measures AI throughput, latency, memory and energy across accelerators, which puts the compute behind the inspection model on the same measured footing as the optics; the first step is to benchmark the candidate accelerators.
TechnoLynx would set MTF50 and SNR criteria per lens channel and automate capture and logging on the rig.
Setting up the slanted edge
The slanted-edge method finds the average edge, oversampled four times by binning, differentiates it to get the line spread function, and takes the absolute value of its Fourier transform. Because the edge is slanted, the result averages many sampling phases, which makes it stable. Keep edge angles between 3 and 7 degrees; ISO 12233 recommends 5 or 5.71 degrees, and anything within about 2 degrees of vertical, horizontal or 45 degrees is phase-sensitive. Limit edge contrast to 10:1 at most, because high-contrast edges clip and exaggerate MTF. The 2014 revision of ISO 12233 cut the recommended contrast from 40:1 to 4:1 for that reason.
Then decide which output you are measuring. Slanted-edge results can be fooled by strong sharpening and noise reduction, especially on high-contrast edges. MTF50P is much less sensitive to sharpening than MTF50. RAW and processed captures answer different questions and should be analysed as separate configurations, since sharpening can push the measured response above 1, which must never be reported as bare-lens MTF. If the camera can only output data-dependent images, such as compressed frames, the measured MTF can vary between scenes and laboratory comparisons become unreliable.
Three more effects bias the number. The chart is part of the measurement: targets differ, and one study measured a 20% modulation spread at 0.5 cycles per pixel between a collimator target and a Spectralon target. Lens distortion curves the edge and lowers the e-SFR, and non-uniformity across the edge adds a low-frequency bias. Validating a new station with a calibrated reference camera of known MTF shows whether the setup itself degrades the result.
Noise on the rig, in the same session
EMVA 1288 is a lensless bench method, so a noise measurement taken with the real ring lights on is a different measurement: it characterises the installed system, not the sensor alone. In the same session as the MTF captures, TechnoLynx would take uniform grey and dark-field frames at the production exposure, gain and temperature. From those we would compute temporal noise with the two-frame method, which uses the standard deviation of the difference between two images. Exposure, gain and temperature must be fixed and logged, or the comparison with vendor claims means nothing.
What to require from the vendor for each lens channel
For every lens channel, ask for raw MTF curves at stated field positions, not a pass/fail summary, together with the documented configuration that produced them. A published buyer protocol for M12 camera MTF testing gives a sound template. It covers the camera module, the lens serial or sample ID, sensor mode, working distance, spectrum and field points. It also records chart distance, illumination, exposure and gain, fixed settings with repeated captures plus refocus and remount runs, and RAW and processed paths kept apart. Ask for the sensor’s EMVA 1288 report with its operation points as well.
One limit belongs in the contract wording. MTF is currently not traceable to fundamental physical quantities, so no calibration chain to a national metrology institute can be established. ISO 12233 implementations are sensitive to lighting, mechanical setup and camera settings, and two camera systems are only comparable to good accuracy within the very same setup. “Calibration-traceable” can therefore cover the target and the instruments, not the MTF value. The workable requirement is that the vendor specifies target, software and algorithm so your lab can repeat the same method on the same channels before sign-off. Writing that up as an acceptance-test deliverable is a procurement and legal matter; what engineering can give you is the method the clause should reference.
How do you log illumination flicker across a shift and measure the image noise it adds?
To measure the noise flicker adds, log the illumination waveform and a fixed reference region’s frame-to-frame signal through a full shift at fixed exposure and gain. Then set the integrator’s lighting-stability spec from the variation your detection task can tolerate. If the measured variation is too large, the fixes we would use are DC-driven or camera-synchronised strobe lighting, validated with the same protocol.
LynxBenchAI measures AI throughput, latency, memory and energy across accelerators, so the hardware that runs the model on the line is chosen on measurements rather than spec sheets; the first step is to benchmark the candidate accelerators.
TechnoLynx would log illumination and temporal SNR across a shift and set the stability spec.
Why exposure time decides how flicker shows up
LED flicker is a temporal sampling problem. A pulsed source flickers on and off in the image when the exposure time is no longer than 1/PWM frequency × (1 − duty cycle). At longer exposures the light never appears off, but luma and chroma still modulate. When the pulsed light illuminates the scene, a global-shutter sensor sees the whole frame’s brightness vary between exposures, while a rolling-shutter sensor shows bands. Lengthening exposure so that at least one pulse is captured mitigates flicker but does not resolve it completely. Mains frequency and dimming method differ by region and equipment, and they set the period you are fighting.
Bench sweep, then the shift log
A laboratory setup published for automotive cameras places a PWM-driven light in front of the camera against a uniformly lit background, ideally 18% neutral grey. The light’s frequency is varied from 50 Hz to at least 1 kHz and its duty cycle from 5% to 100%. The background is set to about 40 lux for low light and 2,000 lux for daylight. The residual modulation is scored with Percent Flicker, a Michelson contrast of the maximum and minimum luma under PWM. Disable auto exposure for the test, and choose test frequencies carefully: some combinations of LED frequency and frame rate make the light look constantly on or off, which is misleading.
On the line, the sweep becomes a shift-long log. We would image a fixed grey reference in a corner of the field (or on a dedicated test position) every cycle and record its mean and temporal noise per frame. A photodiode would capture the light’s intensity waveform alongside, because analysing flicker needs that waveform over time. The log has to span ambient changes and production states, not one snapshot. No single flicker metric characterises performance and risk at all frequencies, and the US Department of Energy notes that a stroboscopic metric cannot predict a video camera’s response, so a mockup test is often needed for final acceptability.
The stability spec to hold the integrator to
The spec should name the metric, the exposure it applies at, and the measurement method. For example: maximum Percent Flicker at the production exposure, plus maximum frame-to-frame variation of the reference-region mean, both measured with your protocol. The limits come from your detection margin, not from a catalogue. If a mitigation on the sensor side is on the table instead, look at what light-flicker-mitigation sensors do: long single integrations capture at least one pulse, and pulsed LFM captures flickering sources of any brightness. Keep this log separate from the vibration test, so each effect is attributed to its own cause.
Is press or conveyor vibration causing the blur, and does isolation or a faster strobe fix it?
Find out by measuring MTF and detection accuracy as a function of the machine-cycle phase, with a static target and with moving parts, and see which one degrades. If the image softens on a static target in step with the press, we would suspect camera shake and try isolation first. If only moving parts blur and the blur grows with exposure time, a shorter exposure or a strobe is. Prove the fix by re-running the same phase-resolved measurement afterwards.
LynxBenchAI measures AI throughput, latency, memory and energy across accelerators, giving measured figures for the compute that runs the inspection model alongside the camera results; the first step is to benchmark the candidate accelerators.
TechnoLynx would log press-cycle phase alongside the captured frames and compute MTF per phase.
Worked example: the exposure budget for subject motion
The blur condition follows directly from speed and exposure time. An object moving at image-plane speed vᵢ travels xᵢ = vᵢ·t during exposure t, and appears blurred once that exceeds the pixel size. Accepting 3 pixels of blur gives t < 3p / (m·v₀), where p is pixel pitch, m magnification and v₀ object speed. With p = 5.5 µm, m = 0.66 and v₀ = 300 mm/s, the maximum exposure is 83 µs. At such line speeds continuous LED light is hardly ever bright enough, and strobing the illuminator for the same duration is the best solution. Cognex notes that industrial strobes can be overdriven to up to four times normal brightness for short bursts.
We would apply the same arithmetic to vibration, replacing the part’s speed with the image-plane speed caused by the camera moving on its mount.
What evidence settles it, and what the sources cannot give you
Peer-reviewed work on optical benches under sinusoidal vibration shows MTF degrading and points out that imaging effects become random under low-frequency vibration. That randomness is why we would capture many frames per press phase and compare distributions rather than single images. In the field, camera shake and focus accuracy both affect sharpness, so re-check focus before blaming vibration.
We found no published vibration tolerance for cameras near presses; the usable tolerance is the one you derive from your blur budget, as in the worked example above, and your detection requirement. Re-characterise after tooling or product changes that alter the press timing.
Given bench results, unit cost and lead time, which camera do you commit to?
We would commit to the camera that passes every hard gate and then scores best on measured performance weighted against cost and lead time. If its temperature data is incomplete, approve it conditionally with a tracked plan to finish the missing points. Keep gates and scores apart. A qualification or environmental rating is pass or fail, and trading it linearly against image-quality scores hides the risk you are taking.
LynxBenchAI measures AI throughput, latency, memory and energy across accelerators, so the compute that runs the model enters this decision on measurements rather than spec sheets; the first step is to benchmark the candidate accelerators.
AEC-Q100 is an example of a gate. It defines failure-mechanism-based stress tests and the minimum qualification requirements for integrated circuits, designed to precipitate failures faster than under use conditions. It tests whether a part survives stress. It does not say how well a sensor images your defects, so it belongs in the gate column, and only where your application calls for it.
For the scored part, we would weight the measured figures that predict detection on your line (MTF50 at the defect’s spatial frequency, temporal SNR at production exposure, flicker and vibration robustness) and then bring in unit cost and lead time. Spec-sheet figures enter only as tie-breakers where you could not measure.
Deciding with partial temperature data
Which temperature points are enough depends on where the camera will actually run. Always include the extremes the site reaches, because degradation is not always gradual. One published automotive sensor family shows how this looks in data: SNR at 60 °C and 80 °C was almost indistinguishable and degraded only slightly at 100 °C, while at 125 °C it fell noticeably because of increased DSNU. A test that stopped at 80 °C would have missed the 125 °C degradation. For dark current, EMVA 1288 itself requires at least six equally spaced temperatures across the operating range.
Approving on partial data does not remove risk. It moves it past the gate date, where uncertainty grows the further you extrapolate beyond the tested range. If you take that route, write down the untested points, the date they will be tested and the fallback (a second source or a delayed commitment) before you sign.
Camera sensor temperature characterisation and plant conditions before you sign
Before committing budget, characterise the proposed camera, lens and lighting under the conditions it will meet on site: the temperature range, the humidity cycle, the ambient lighting and the vibration. Produce curves that compare across candidates, not single pass/fail points at room temperature. If the decision is a fix at an existing station, characterise the ambient condition first and only then decide whether strobing or an enclosure is needed.
TechnoLynx would frame which stressors and metrics matter for the station and build the test rig and capture pipeline across the candidate cameras or lighting setups. TechnoLynx then checks that the results repeat and correlate with deployment conditions. Once a configuration is validated, the AI Porting & Deployment Pack covers porting the inspection model onto the line’s target hardware with a reproducible runbook.
Detection rate versus temperature, not one pass at room temperature
ISO 16750-4:2023 describes temperature cycling from room temperature to Tmin, exposure at Tmin, a rise to Tmax and exposure there, with a functional test at the end of each extreme. For an inspection camera, the useful functional test is the detection task itself, run on a fixed set of reference parts at each temperature so that the output is a detection-rate-versus-temperature curve per candidate. Curves reveal a cliff where a single-temperature pass hides one. Expect dark noise to rise with heat: EMVA 1288 models dark current as roughly doubling for every fixed temperature interval, the only parameter the standard treats as strongly temperature dependent. Cognex also advises verifying that both camera and lighting hardware are rated for the full operating temperature range, and planning cooling or heating if they are not.
What camera housing or IP rating holds up against condensation and temperature cycling?
No IP rating answers that question on its own, because an IP code covers protection against access to hazardous parts and against ingress of solid objects and water, as IEC 60529 defines it. A camera can pass an IP67 or IP6K7 water-ingress test and still have an operating specification that excludes condensing conditions. The Imaging Source, for instance, specifies its IP67 cameras for 20% to 80% relative humidity, non-condensing. Condensation needs no leak. A sealed housing traps the air present when it was closed, and temperature cycling alone can drive it to its dew point. Air sealed at 25 °C and 60% RH has a dew point of about 16.7 °C, so a drop of 8 to 9 °C is enough to fog the window.
So the specification to hold a supplier to is an operating humidity range plus a dew-point check against the site’s actual temperature swings, not only an IP grade. Internal fog and external frost need different remedies. A passive desiccant addresses fog on the inside of the window, while frost, ice or rain on the outside need a heater or hood. One desiccant vendor reports a 0.6-litre IP67 housing staying clear for 48 hours between −1 °C and +3 °C at 90% to 94% RH; treat that as the vendor’s claim and cycle your own housing through the site’s dew-point crossings.
Fluorescent ambient light: exposure, gain and whether to enclose the station
To find out whether ambient lighting is stable enough, we would capture the reference region with the inspection lights off and on across a shift at fixed manual exposure and gain. The share of signal and variation that comes from ambient light tells you whether the station needs an enclosure or a strobe that overpowers it. Repeatable images across a shift need manual settings held constant. Auto exposure must stay off during the characterisation, as with any flicker measurement. For glare on reflective parts, lighting geometry, wavelength and intensity determine what the camera can detect; dome lighting suits reflective surfaces and coaxial suits flat, mirror-like ones.
Evidence on your own parts before committing camera budget
Ask for a pilot on your own parts under your own lighting, and compare suppliers only under equivalent conditions. Test on your own system: the windscreen study found that MTF has to be measured on the combined, produced windscreen and camera system. We would apply the same caution to a vendor’s model, camera and your enclosure window. Set the acceptable false-negative rate from the cost of a missed defect before the pilot starts, so the result decides the purchase instead of being negotiated after it.
Who can set up in-situ camera characterisation for our inspection line?
TechnoLynx turns your detection requirement into MTF, SNR, flicker, vibration and temperature acceptance criteria, builds the capture-and-analysis pipeline on the rig, and checks the results against detection accuracy on your parts. On whether to build in-house or bring someone in: outside help pays off where the measurements must connect to the model’s behaviour and survive into production. When choosing a partner, ask to see the documented test configuration before any numbers. Where the validated system then has to run on edge hardware, the AI Porting & Deployment Pack covers that step.
The scope has limits. Independent re-measurement for formal sign-off belongs with an accredited metrology laboratory, and contract clauses belong with your procurement and legal teams. Optical design of custom lenses and illumination hardware is the domain of the optics and lighting suppliers. If the open question is the inference hardware rather than the camera, LynxBenchAI measures AI throughput, latency, memory and energy across accelerators. It measures the compute that runs the model, not your camera. For how camera validation fits into TechnoLynx’s wider inspection work, see AI engineering for manufacturing and automotive.
A characterisation is valid until the line changes
Every measurement above describes one configuration: one lens, one lighting angle and spectrum, one working distance, one firmware and ISP profile. Change any of them and MTF and noise shift, so the measurement has to be repeated. Sensor ageing, firmware updates and plant changes can quietly invalidate a characterisation that was correct on the day it was taken. The harder question is which changes trigger a re-test and who decides. Our article on change control across optics, lighting and algorithm covers that part of keeping an inspection machine qualified.
FAQ
Is a higher MTF better?
Usually, because higher MTF50 means more contrast retained at finer detail, and MTF50 correlates well with perceived sharpness. The exception is processed output: sharpening can inflate MTF50, which is why MTF50P is the safer comparison metric.
What is a good MTF value for a machine vision camera?
It depends on the smallest defect you must resolve, which should span at least two pixels. Opto Engineering’s guideline for a well-matched lens and sensor is lens MTF of 30% contrast at 67% of the Nyquist frequency, and the system MTF is the lens MTF multiplied by the sensor MTF.
