CCD vs CMOS image sensors

Both sensor types collect photoelectrons in a two-dimensional array of silicon pixels, and both are limited to wavelengths shorter than about 1.1 µm, where photons exceed silicon's 1.12 eV bandgap. They differ in how the collected charge is read out. In a charge-coupled device (CCD), clocked gate voltages move each pixel's charge packet row by row into a serial register and then pixel by pixel to a single output amplifier, so every pixel is measured by the same amplifier. In a CMOS active-pixel sensor, each pixel has its own photodiode and source-follower transistor, and the signal is digitized in parallel by an analog-to-digital converter at the foot of each column. The CCD was invented by Boyle and Smith at Bell Labs in 1969; CMOS sensors displaced it in consumer cameras during the 2000s and, in the form of scientific CMOS (sCMOS), in much of microscopy and laboratory imaging during the 2010s.

Readout, speed and noise

Serial readout makes a CCD slow. A 2048 × 2048 sensor (4.2 megapixels) read through one output at 10 MHz takes 0.42 s per frame, and the read noise of the output amplifier rises with pixel rate; slow-scan scientific CCDs reach a few electrons rms by reading at well below 1 MHz. Column-parallel readout lets a CMOS sensor read the same number of pixels at 100 frames per second or more, and because each column ADC runs slowly, read noise stays low: 1–2 electrons rms is typical for sCMOS.

A CCD transfers charge thousands of times before it is measured, so the charge transfer efficiency (CTE) matters. For a pixel 2048 transfers from the output,

$$\text{retained fraction} = \mathrm{CTE}^{2048},$$

which is 0.980 for CTE = 0.99999 and 0.998 for CTE = 0.999999. Radiation damage lowers CTE, which is a concern for CCDs in space.

Signal-to-noise ratio

For a signal of $S$ photoelectrons, dark charge $D$ and read noise $\sigma_r$, all in electrons,

$$\mathrm{SNR} = \frac{S}{\sqrt{S + D + \sigma_r^2}},$$

where the first term under the root is shot noise. With $S = 100$ e⁻ and negligible dark charge, a read noise of 1.5 e⁻ gives SNR = 9.9, close to the shot-noise limit of 10, while 8 e⁻ gives 7.8. At $S = 10$ e⁻ the same two read noises give 2.9 and 1.2: at low light, read noise dominates the noise budget. Electron-multiplying CCDs (EMCCDs) address this by amplifying the charge before the output, reaching effective read noise below one electron at the cost of an excess noise factor of about √2, which halves the effective quantum efficiency.

For long exposures the balance shifts. Deep-cooled CCDs have very low dark current, a few electrons per pixel per hour or less in astronomical sensors cooled to about −100 °C, and a single output amplifier means every pixel has the same gain and offset. This uniformity and the low dark charge keep CCDs in astronomy, spectroscopy and some X-ray instruments.

Quantum efficiency and shutter

Back-illuminated sensors of either type, thinned so light enters from the side without metal wiring, reach peak quantum efficiencies above 90% in the visible; front-illuminated CMOS pixels lose light to the wiring and rely on microlenses. The near-infrared response of both falls off toward 1.1 µm because silicon's absorption length becomes longer than the pixel depth. For wavelengths beyond that, an InGaAs SWIR camera is needed.

Interline and frame-transfer CCDs move the whole frame into a masked storage area at once, which acts as a global shutter; full-frame CCDs need a mechanical shutter. Many CMOS sensors use a rolling shutter, exposing and reading row by row, so moving objects or pulsed illumination can appear skewed or banded; global-shutter CMOS pixels exist at some cost in noise or fill factor.

Pitfalls

CMOS sensors show fixed-pattern noise from column amplifier and pixel offset differences, and some pixels exhibit random telegraph noise that appears as flickering hot pixels; calibration frames correct the fixed pattern, and median filtering or pixel masking suppresses the telegraph pixels. CCDs bloom when a saturated pixel spills charge along the column, and smear appears when light falls during transfer. For laser work, a beam profiler built on either sensor needs attenuation to avoid saturation, and a rolling shutter interacts badly with pulsed lasers at low repetition rates. Many datasheets quote QE, read noise and dark current under the EMVA 1288 standard, which makes cameras comparable under stated conditions.

Common questions

Is CMOS better than CCD?

For most imaging today, yes: CMOS is faster, cheaper, lower power and, in scientific versions, has lower read noise at useful frame rates. CCDs keep an advantage for very long, deeply cooled exposures with minimal dark current and high uniformity, and EMCCDs for photon-starved, single-photon-level imaging.

What is scientific CMOS?

A CMOS sensor designed for quantitative imaging, with read noise around 1–2 electrons, high dynamic range from dual-gain readout, and frame rates of tens to hundreds of frames per second at several megapixels.

Why can't silicon cameras see 1550 nm light?

A 1550 nm photon carries 0.80 eV, below silicon's 1.12 eV bandgap, so it is not absorbed. Silicon sensors cut off near 1.1 µm, and telecom-band imaging uses InGaAs arrays.

References: J. R. Janesick, Scientific Charge-Coupled Devices (SPIE Press, 2001); J. R. Janesick, Photon Transfer (SPIE Press, 2007); W. S. Boyle, G. E. Smith, Bell Syst. Tech. J. 49, 587 (1970); EMVA Standard 1288, Standard for Characterization of Image Sensors and Cameras.