Read Noise

Read noise is the uncertainty the electronics add each time a pixel is read. It is independent of signal and independent of exposure: a pixel read once carries it once, whether it holds three electrons or thirty thousand.

Sub-electron read noise is the headline specification on premium scientific cameras, and here is what it is worth at a typical working signal. A pixel collecting 400 photoelectrons carries √400 = 20 electrons of shot noise. Add 0.7 electrons of read noise in quadrature and the total is √(400 + 0.49) = 20.012. Use a camera at 0.3 instead and the total is 20.002. The difference in SNR is one part in two thousand.

Not a small advantage — an unmeasurable one. At 400 photons the read-noise specification is not a tiebreaker, it is not a factor, and the premium buys nothing you could detect in a hundred frames.

There are exactly two regimes where read noise earns its price, and both are identifiable before you buy anything.

Very low signal. Below roughly five detected photons per pixel, shot noise is smaller than the read noise and the detector is genuinely the limit. Below about 0.3 electrons the output histogram separates into discrete photon peaks and true counting becomes possible. This is single-molecule territory, and there the difference between 0.7 and 0.3 is decisive rather than decorative.

Subdivided exposure. Read noise is paid per read, so N frames pay it N times. Splitting a 19-electron acquisition into ten frames costs about 10% of SNR at 0.7 electrons and about 2% at 0.3. If your experiment must be sliced finely — drift correction, registration, bleach tracking — sub-electron performance is what makes the slicing nearly free.

A CCD reads its whole array through one amplifier, so read noise is genuinely a single value. A CMOS sensor gives every pixel its own amplifier, which is what makes it fast — and means read noise varies from pixel to pixel across the array.

The quoted figure is therefore a statistic, usually the median, and a median hides a tail. A sensor advertised at 1.0 electron median may have a small population above 3 electrons, and those pixels will produce convincing false positives in any low-light detection task — they look exactly like faint puncta. The RMS figure is higher than the median and is the more honest single number; better still is the full histogram, which good manufacturers publish.

The practical defense is a dark-frame map: characterize your own array once, identify the noisy pixels, and treat them as known bad rather than as data.

You need not take the datasheet's word for it. Collect a hundred dark frames at the exposure and readout mode you actually use. For each pixel, the mean across the stack is its offset; the standard deviation across the stack is its read noise in counts. Multiply by the conversion gain and you have electrons.

Worked through: a hundred frames at 100 ms, per-pixel temporal standard deviation averaging 46.7 counts, conversion gain 0.015 electrons per count. Read noise = 46.7 × 0.015 = 0.70 electrons, confirming the specified sub-electron mode in the mode you are running rather than the one the brochure was measured in. Keep the spatial spread of those per-pixel values too: that is your noise histogram, and its tail is the thing the median hid.