Quantum efficiency curve data: how to read a camera QE plot

You have a monochrome camera, its manual shows a quantum efficiency (QE) plot, and you want the numbers behind it — a table you can multiply by photon flux to predict signal levels. Most camera makers publish the curve but not the table, and the reason is not secrecy. The values come from one test bench and can move 20-30% on another, so they are comparison data, not calibration data. What follows covers what the curve measures, where its numbers come from, how to read one to choose a sensor or a light wavelength, and what to do when a calculation genuinely needs per-wavelength values.

What does a quantum efficiency curve actually measure?

It plots the fraction of incoming photons that the sensor converts into electrons, wavelength by wavelength: a QE of 0.8 at 500 nm means eight of every ten photons reaching a pixel produce a photoelectron. The vertical axis runs from 0 to 1 and the horizontal axis from near ultraviolet into the near infrared; for silicon sensors the plotted span typically covers 400-1000 nm, with real sensitivity stretching from about 350 to 1050 nm. Four quantities tie the curve to a signal you can predict:

QE(λ) is electrons out over photons in at wavelength λ, E_ph is the energy of one photon, and S is the charge a pixel of area A collects from irradiance Φ over exposure time t. R is the responsivity in amps per watt that radiometry works with. You can measure or look up every term except QE; the curve's one job is to supply QE at your wavelength.

Why won't the datasheet give QE values as a table?

Because the curve is one measurement on one bench: the same camera can come out 20-30% higher or lower on another EMVA 1288 setup, and a comma-separated table would invite calculations the data cannot support. EMVA 1288 anchors camera sensitivity with monochromatic light at the wavelength where QE is maximal, holding the illumination band within about 50 nm. The full sweep printed in a manual is a separate relative spectral-response run on the maker's own equipment — reference detector, monochromator, lamp stability, sensor cover glass — and each of those contributes a few percent of error. Combined, two competent laboratories routinely land 20-30% apart.

Makers therefore publish the shape and release the underlying values through their support channel, with the uncertainty restated. Raise a ticket with the exact model and serial number, say the numbers will be used comparatively, and you will generally receive the table. If you are deciding between camera models, ask through our contact page and we can help you put that request to the maker and read what comes back. For the broader skill of judging cameras from standardized measurements, start with reading EMVA 1288 performance data.

What does a typical monochrome CMOS QE curve look like?

Expect roughly 0.40-0.55 at 400 nm, a broad peak of 0.70-0.85 between 500 and 600 nm, a slow fall through red, and a steep drop toward zero approaching 1000 nm where silicon stops absorbing. The bands below are typical for current CMOS sensors, not one product's specification; a widely quoted monochrome example reads 0.5 at 400 nm and 0.8 at 500 nm, which sits inside them.

A color camera prints three curves instead of one. Each pixel sits under a red, green or blue section of the mosaic, and those filters absorb before the silicon converts, so a color sensor typically delivers 2-3 times fewer electrons from white light than the monochrome build of the same silicon. The monochrome and color versions of one sensor family make the difference visible:

All 2 share: Sensor type CMOS, rolling shutter; Sensor model Sony® IMX183; Pixel size 2.4 μm × 2.4 μm; Sensor size 1"; Resolution 5472 × 3648; Dynamic range 65.5 dB; SNR 41.5 dB; Gain 0 dB to 24 dB. Values as printed in each product's data sheet; open a part for its drawing and full specifications.

How do you compare two cameras from their QE curves?

Read both curves at the wavelength of your light, on plots from the same maker and test setup, and treat any gap smaller than about 20% as measurement spread rather than a real difference. Three checks keep the comparison honest. First, know whether the vertical axis is absolute (0-1) or normalized to the peak at 1.0 — a normalized plot tells you the shape, never the value to put in a formula. Second, compare curves from the same bench: differences beyond 20% at your wavelength mean something between one maker's cameras, and between different makers treat 30% or more as the threshold. Third, compare like with like: mono against mono, or the same color channel against itself.

For narrowband or filtered work the monochrome camera wins on physics, not preference — the reasoning is in monochrome camera with an optical filter. The curve also tells you which LED color to buy, which is where light color and monochrome camera contrast starts. A monochrome model such as the MV-CS200-10UM gives you the single broadband curve this kind of comparison needs.

The same reading discipline applies to lens performance curves: one plot, one bench, compare at your working point. The camera finder filters by spectrum and shutter before you open a single datasheet.

Can you use QE values for absolute exposure calculations?

Not to datasheet precision: a 20-30% uncertainty in the QE value passes straight through to the predicted signal, so use the curve for ratios between wavelengths and cameras, and set the final exposure on the bench. Ratios survive because the bench errors cancel when both wavelengths ride the same curve; absolute electron counts do not, because every term inherits the setup error.

The 2:1 ratio between the two wavelengths is trustworthy; the absolute 2.4×10⁴ electrons is not. If the true QE at 500 nm is 0.6 rather than 0.8 — squarely inside the 20-30% band — the same exposure yields 1.8×10⁴ electrons, a 25% shortfall you would only find on the machine. And if the 400 nm value is really 0.35, the exposure that matches the 500 nm channel grows from 2 ms to about 2.9 ms.

When your light's output is specified in lux or lumens rather than watts, convert first: radiometric versus photometric light specification covers the units.

How do you measure a camera's relative spectral response yourself?

Fix the camera and its exposure, illuminate a stable white target through a set of narrowband filters, and record the mean gray value per filter — the ratios between filters give you relative spectral response on your own bench. A workable procedure:

1. Mount a current-driven light source and a diffuse white target, blackout ambient light, and lock gain and exposure.
2. Set the exposure so the brightest filter lands near 70% of full scale, below saturation with headroom for drift.
3. For each filter, record the mean gray value in a central region of interest and subtract a dark frame taken at the same exposure.
4. Plot dark-corrected gray value against each filter's center wavelength.
5. Compare two cameras through the same filters on the same bench: the lamp's own spectrum cancels. Comparing wavelengths on one camera does not cancel it — for that you need a calibrated reference detector.

Ten-nanometer interference filters at 405, 470, 525, 590, 625, 660, 730, 850 and 940 nm cover the band most machine-vision lights use. If the goal is separating materials rather than characterizing the camera, plan it as a validation experiment, the way spectral imaging material classification describes.

When should you ignore the QE curve and test instead?

At the edges and behind the glass: below about 450 nm, above about 900 nm, and behind any window, filter or lens coating that reshapes the spectrum, the curve is steepest and a 20-30% error flips a marginal pass into a fail. Three cases account for most surprises. First, 405 nm and shorter: QE is falling fast there and sensor-stack filters differ between models, which is why the practice of testing UV contrast before you buy exists. Second, 850 nm and 940 nm: the tail drops several percentage points per 10 nm of wavelength there, and protective windows and lens coatings are rarely flat past 850 nm. Third, beyond about 1000 nm silicon is finished: the curve ends and the application needs a SWIR sensor, not a longer exposure. Board-level cameras such as the LS-BCM-0003 meet the same spectral edges inside embedded housings, where the window over the sensor is often the one part nobody measured.

What should you check before trusting a QE plot?

Run five checks in order: the axis units, the channel, the sensor variant, the test setup, and where your wavelength sits on the curve. Each takes seconds and each prevents a specific mistake.

• Absolute or normalized axis — a 0-1 axis feeds the formula above; a peak-normalized axis only compares shapes.
• Mono or color channel — mono is one curve; color is three, and the green curve is not the sensitivity of the camera.
• Sensor variant and cover glass — the same sensor name can carry different filter stacks; confirm the curve matches the camera model.
• Same-bench comparison — gaps under 20% between makers are noise; gaps over 30% are decisions.
• Position of your wavelength — on the flat top between 500 and 600 nm the curve forgives; on the slopes it does not.

Then decide. Comparing sensors under a 525-625 nm light from one maker: trust the curves. Crossing makers, or working at 405 nm or 850/940 nm: put the candidate on your bench and measure. Predicting absolute electrons: measure, or keep the 1.5-2× margin. Start candidate lists in area scan cameras or the camera configurator, and read further in the cameras topic.