Quantum efficiency at 525 nm and dark noise in electrons are the two most decision-relevant sensor specs on a machine-vision camera datasheet, and the EMVA 1288 standard gives a common basis for comparing them.
Typical 525 nm performance bands from published measurements are 50-70% QE for front-illuminated CMOS, 60-80% for back-illuminated CMOS, and 80-95% for back-illuminated scientific CMOS (sCMOS) and BSI CCD; cooled sensors drop temporal dark noise to 0.7-2 e-, while uncooled industrial cameras run 5-15 e- [S2][S3][S4].
QE at 525 nm: where each sensor type lands
525 nm is the green wavelength most white-LED machine-vision lines and many laser-line applications use, so QE at that wavelength is often quoted separately from peak QE on the spectral curve [S1][S5]. A front-illuminated (FI) CMOS with a standard microlens typically lands at 50-70% QE at 525 nm, because the polysilicon gate stack and on-pixel transistors block or absorb a fraction of the photons before they reach the photodiode [S3][S5]. Back-illuminated (BI / BSI) sensors flip the die so light hits the photodiode first, lifting 525 nm QE into the 60-80% range for industrial BSI parts and into the 80-90% range for sCMOS-grade BSI, with certain BSI CCDs reaching 80-95% peak and still scoring 75-85% at 525 nm [S2][S3]. The improvement comes from a thinner active silicon layer, anti-reflection coating optimisation, and in some designs a nanostructured surface that reduces reflection at a specific wavelength [S2].
It is common for QE to drop sharply between 400 nm and 450 nm on standard silicon sensors, then plateau across the green band (500-560 nm) before falling off again past 650 nm, which is why a sensor that performs well at 525 nm can underperform the same vendor's BSI sibling in the blue [S1][S5]. A reading taken at 525 nm is therefore more representative of a green-LED production line than a peak-QE number printed on a marketing sheet [S4].
Dark noise in electrons: cooled vs uncooled sensors
Dark noise is the standard deviation of pixel signal with no light, expressed in electrons (e-), and is dominated by thermally generated carriers and by the read-noise floor of the on-chip amplifier [S4][S7][S9]. Uncooled industrial CMOS cameras typically quote 5-15 e- RMS temporal dark noise at room temperature, while scientific and interline-transfer CCDs and sCMOS sensors drop to 1-3 e- at ambient conditions and 0.7-2 e- when thermoelectrically (TEC) cooled to -10 °C to -25 °C [S3][S7][S10]. Dark current roughly halves for every 7-10 °C drop in sensor temperature, so a sensor at 30 dB above ambient thermal noise can fall below 1 e- per pixel per second of integration when cooled [S9][S10].
Dark noise is independent of QE, which is why EMVA 1288 and Adimec both recommend the ratio QE / read noise as a single-figure sensitivity metric, rather than quoting either parameter alone [S4][S5]. For a pick-and-place line in a temperature-stable cabinet, an uncooled 8 e- sensor with 60% QE at 525 nm may be perfectly adequate; for fluorescence microscopy or low-light metrology, a 1.2 e- cooled sCMOS with 88% QE at 525 nm is the correct choice even at 5-10x the unit cost [S2][S3][S7].
Selection criteria: 525 nm QE, dark noise, and what they cost
Five decision criteria tend to dominate the camera spec sheet on a 2026-vintage industrial vision project: 525 nm QE, dark noise (e-), full-well capacity (ke-), pixel pitch (µm), and interface throughput (frames per second at the chosen region of interest) [S3][S4]. On those criteria, the four common sensor families line up as follows: front-illuminated CMOS gives 50-70% QE at 525 nm and 5-15 e- dark noise but the highest frame rates and lowest cost; back-illuminated CMOS gives 60-80% QE at 525 nm with similar dark noise to FI CMOS and slightly higher cost; scientific CMOS (BSI, often cooled) gives 80-90% QE at 525 nm and 0.7-2 e- dark noise at the price of lower full-well capacity and higher system cost; BSI CCD gives 80-95% QE at 525 nm and 1-3 e- dark noise but lower frame rates, higher power, and increasingly limited supply [S2][S3][S10].
For a battery-inspection line lit by 520 nm green LEDs, the marginal benefit of moving from 65% to 88% QE at 525 nm is roughly a 35% reduction in required exposure time at matched SNR, which directly raises line throughput; for a barcode-reading station lit by a 660 nm red laser, the same 65% to 88% jump may collapse to under 10% because both sensors are rolling off the red side of the curve, so the QE gain is wavelength-dependent [S1][S5]. Selecting on the ratio of QE to dark noise rather than either number alone avoids overspending on dark-noise performance the application never uses [S4][S5].
EMVA 1288, vendor datasheets, and how to read the curve
The EMVA 1288 standard is the European Machine Vision Association framework for measuring and reporting QE, dark noise, saturation capacity, and dynamic range under controlled monochromatic illumination, and most reputable machine-vision vendors publish EMVA 1288 data sheets alongside marketing QE curves [S2][S4][S5]. A practical reading rule: take the QE value at the exact wavelength of the dominant illumination source, not the peak, and read dark noise at the same sensor temperature and gain that the application will actually use, because the two drift together [S1][S4][S7]. For multi-spectral or white-light applications, the 2025 AQE method proposed by Deng et al. offers a broadband-averaged QE figure that better predicts real-world response than a peak-QE number [S2].
When comparing two cameras, normalising the SNR-limited exposure at the same scene irradiance is the most useful check: a higher-QE but noisier sensor can lose to a lower-QE but quieter sensor at very low light, while at moderate light the QE winner almost always wins on speed [S4][S5][S7]. Verifying the sensor-temperature spec in the datasheet is equally important, since a 1 e- figure is only valid at the listed cooling setpoint, not at the ambient 25-35 °C found inside most enclosures [S9][S10]. See our industrial borescope inspection guide for how low-light detector selection principles translate to confined-space visual inspection, and revisit industrial adhesive selection for an unrelated but spec-driven component decision that uses a similar datasheet-reading discipline.
Failure modes and known limitations
The two failure modes that catch engineers out are: (1) trusting a peak-QE number on the marketing curve for a wavelength the application does not use, which routinely inflates expected sensitivity by 15-30%; and (2) ignoring that dark noise rises with integration time, so a sensor specified at 1 e- read noise but driven at 1 s exposures may see effective noise climb to several electrons once dark current is integrated [S1][S4][S9]. A third, less obvious mode is that global-shutter pixels, often required for moving-line inspection, add an extra transistor that reduces fill factor and trims QE by 3-8 percentage points compared to a rolling-shutter variant of the same die, so shutter choice is a real QE trade-off [S3][S5].
Cooled sensors also introduce condensation risk on the window when the housing is opened in a humid plant, which is why most industrial sCMOS products integrate a sealed dry-gas-purged window rather than a user-serviceable one [S9][S10]. For a relevant cross-reference on matching sensor specs to system architecture, our GigE Vision vs USB3 Vision vs Camera Link spec sheet walks through the interface-throughput side of the same selection problem.
Who benefits from high 525 nm QE and who does not
The buyers who benefit most from 80%+ QE at 525 nm are fluorescence-imaging lab groups, low-light semiconductor inspection, biomedical sorting, and astronomy-adjacent metrology, where every photon counts and exposure budgets are tight [S2][S3]. Buyers who rarely need it include outdoor traffic and ANPR systems, high-irradiance web-inspection lines, and most barcode or 2D-code reading stations, all of which are photon-rich and are better served by an uncooled 8 e- sensor with higher full-well capacity and lower cost [S3][S5][S10]. A useful rule: if integration time is below 1 ms at F2.8, dark noise below 5 e- buys almost nothing and the spec can be relaxed in favour of QE, frame rate, or pixel pitch.
For component-level cross-references in other spec-driven decisions, industrial adhesive selection and industrial borescope inspection follow the same read-the-datasheet-not-the-headline discipline. Two verifiable signals to track into 2027: the rollout of stacked BSI global-shutter pixels that aim to recover the 3-8% fill-factor loss noted above, and the wider adoption of EMVA 1288 AQE-style broadband reporting to replace single-wavelength peak-QE marketing numbers [S2][S3].
Spec-level background on the components involved: checking and testing apparatus.