Full-well capacity

Full-well capacity is the maximum charge, counted in electrons, that a pixel can accumulate during an exposure and still report correctly. Above it, extra photoelectrons are lost, spill into neighbors or no longer change the output, and the pixel is saturated. Small CMOS pixels of about 1 µm in consumer cameras hold a few thousand electrons, a 6.5 µm scientific CMOS pixel about 30,000, and the large pixels of slow-scan CCDs and spectroscopic photodiode arrays from 10⁵ electrons to over 10⁷. With the read noise, it sets a sensor's dynamic range, and through shot noise it caps the signal-to-noise ratio a single exposure can reach.

What sets the limit

Charge is stored on a capacitance, so the capacity is the stored charge divided by the electron charge,

$$N_\text{fw} = \frac{C\,\Delta V}{q},$$

where $\Delta V$ is the usable voltage swing. A 10 fF sense node with a 0.5 V swing holds 31,200 electrons, at a conversion gain of $q/C$ = 16 µV per electron. Which element sets the limit depends on the architecture. In a CCD it is the depth of the potential well under the pixel gates, and the summing well and output node must hold any binned charge as well. In a CMOS pixel with a pinned photodiode it is either the photodiode's own capacity or the charge the floating diffusion can take after transfer. The analog-to-digital converter can set a lower ceiling: a 16-bit converter at 0.4 e⁻ per digital number (DN) reaches full scale at 26,214 electrons, so a camera with a 30,000-electron pixel saturates digitally before the pixel does.

Capacity and read noise are linked. A small sense capacitance gives a high conversion gain and low input-referred noise but a small full well; a large capacitance does the reverse. Dual-gain sCMOS readout, described under CCD vs CMOS, reads each pixel through a high-gain and a low-gain channel and combines them to obtain both.

Measurement

Full-well capacity is measured with a photon transfer curve: flat-field frame pairs at increasing exposure, with the variance of their difference plotted against the mean signal. Below saturation the shot-noise variance in DN grows in proportion to the mean, and the slope gives the conversion gain. As pixels start to clip, the variance stops rising and then collapses, because saturated pixels all report the same value. The signal at the variance peak, converted to electrons, is the full well. Many datasheets instead quote a linear full well, the signal at which the response departs from a straight line by a stated amount, which is lower than the absolute saturation level.

Signal-to-noise ratio and exposure

At full well the shot noise is $\sqrt{N_\text{fw}}$, so the best single-exposure signal-to-noise ratio of one pixel is

$$\mathrm{SNR}_\text{max} = \sqrt{N_\text{fw}}.$$

For 30,000 electrons this is 173, or 44.8 dB. Dividing by a read noise of 1.5 e⁻ gives a dynamic range of 20,000, or 86 dB. A higher SNR needs more charge than one well holds, which means averaging frames or binning pixels.

The full well also sets the longest useful integration time for the brightest pixel. A pixel receiving 1 pW at 550 nm absorbs $2.77 \times 10^6$ photons per second; at a quantum efficiency of 0.8 it collects $2.2 \times 10^6$ electrons per second and fills a 30,000-electron well in 13.5 ms. A focused laser beam on a few pixels saturates them far sooner, which is why beam imaging needs strong attenuation, as discussed for the SWIR camera.

Pitfalls

Saturated pixels give no warning in a single frame: the value looks plausible but is too low, and a peak or a spectral line clipped at full well distorts centroids, line ratios and beam widths. In CCDs, charge above full well spills along the column (blooming) and creates bright streaks; anti-blooming drains suppress it but reduce the capacity and can make the response nonlinear well before nominal saturation. The capacity depends on operating settings such as gain mode, clock voltages and, in CMOS sensors, the readout mode, so the value applies to the mode in which it was measured.

Common questions

Is a larger full-well capacity always better?

For bright scenes and high-precision photometry, yes, since it raises the maximum SNR and the dynamic range. For low-light imaging the read noise matters more, and the high conversion gain that gives low read noise usually comes with a smaller well.

Does binning increase full-well capacity?

In a CCD, charge binning adds the charge of several pixels before readout, so the limit becomes the capacity of the serial register and output node, which is usually larger than one pixel's but smaller than the sum. Digital binning after readout adds the pixel values, so the effective capacity is the sum of the individual wells, at the cost of adding each pixel's read noise.

References: J. R. Janesick, Photon Transfer: DN → λ (SPIE Press, 2007); J. R. Janesick, Scientific Charge-Coupled Devices (SPIE Press, 2001); G. C. Holst and T. S. Lomheim, CMOS/CCD Sensors and Camera Systems, 2nd ed. (SPIE Press, 2011).