CCD & CMOS Imaging Sensors — Abridged Guide
Quick-reference guide to imaging sensors — CCD/CMOS architecture, noise, QE, binning, shutters, and selection. For full derivations and worked examples, see the Comprehensive Guide.
Comprehensive Imaging Sensors Guide →
1.Introduction
CCD and CMOS are both silicon-based sensors exploiting the photoelectric effect. They differ in how charge is transferred and read out: CCD transfers charge sequentially to a shared amplifier; CMOS converts charge to voltage at each pixel and reads in parallel.
When comparing sensor specs, focus on four numbers first: read noise (e⁻ rms), quantum efficiency (%), dark current (e⁻/p/s at stated temperature), and frame rate (fps).
2.CCD Architecture
Three CCD variants exist: full-frame (highest fill factor, needs mechanical shutter), frame-transfer (continuous imaging, double silicon area), and interline-transfer (electronic shutter, reduced fill factor). Full-frame and frame-transfer dominate scientific applications.
Full-frame CCDs offer the best QE and lowest noise but require a shutter. If continuous acquisition without a shutter is needed, frame-transfer or interline-transfer architectures are required.
3.CMOS Architecture
The 4-transistor (4T) pinned-photodiode pixel with correlated double sampling (CDS) is the architecture that brought CMOS read noise to sub-electron levels, enabling sCMOS to compete with and surpass CCDs in most imaging applications.
CMOS sensors with column-parallel ADCs read entire rows simultaneously, enabling frame rates 10–100× faster than CCDs at equivalent pixel counts.
4.Noise & SNR
Full SNR
S = signal (e⁻), I_d = dark current (e⁻/p/s), t = integration time (s), σ_r = read noise (e⁻ rms)
Three noise regimes determine sensor choice: read-noise limited (low signal → reduce read noise or use EMCCD), shot-noise limited (moderate signal → ideal regime, collect more photons), dark-noise limited (long integration → cool the sensor).
To quickly estimate if read noise matters: if the signal exceeds ~10× the read noise squared, the measurement is shot-noise limited. For a 3 e⁻ sensor, that threshold is about 90 photoelectrons.
| Noise Source | Depends On | Typical Values | Mitigation |
|---|---|---|---|
| Photon shot noise | Signal level | √S | Collect more photons |
| Read noise | Amplifier, readout speed | 1–15 e⁻ rms | Slow scan, sCMOS, EMCCD |
| Dark noise | Temperature, integration time | √(I_d · t) | Cool the sensor |
| Fixed-pattern noise | Pixel-to-pixel variation | Calibration-dependent | Dark subtraction, flat-field |
5.QE & Spectral Response
Back-illuminated sensors achieve 90–95% peak QE by eliminating wiring obstruction. Front-illuminated sensors with microlenses reach only 55–65%. For any photon-starved application, BSI is the first specification to check.
Deep-depletion CCDs extend useful sensitivity to ~1050 nm by using thick, high-resistivity silicon. They require aggressive cooling due to higher dark current.
| Architecture | Peak QE (%) | UV | NIR | Use |
|---|---|---|---|---|
| FSI CCD | 50–60 | Poor | Moderate | Legacy |
| FSI CMOS + microlens | 55–65 | Poor | Moderate | Consumer, machine vision |
| BSI CCD | 85–95 | Good | Limited | Astronomy, spectroscopy |
| BSI sCMOS | 90–95 | Good | Limited | Microscopy, general scientific |
| Deep-depletion CCD | 85–90 | Moderate | Excellent | NIR spectroscopy |
6.Dynamic Range & Binning
Dynamic Range
On-chip binning (CCD) improves SNR by N in the read-noise-limited regime. Software binning (CMOS) improves by only √N. This is why CCDs persist in spectroscopy: full vertical binning delivers 256× the signal with 1× the read noise.
Dynamic range ≠ bit depth. A 16-bit ADC does not give 16 bits of DR unless N_sat / σ_noise ≥ 65,536.
7.Cooling
Dark Current Halving Rule
Dark current approximately halves for every 5–7 °C of cooling. TE cooling to −70 to −100 °C is sufficient for most scientific imaging. LN₂ cooling is reserved for deep-depletion NIR sensors and ultra-long integrations.
If dark noise is less than read noise, further cooling provides diminishing returns. Compare √(I_d · t) to the read noise spec first.
8.Shutters
Global shutter (all pixels exposed simultaneously) is inherent to most CCDs and essential for fast-moving objects. Rolling shutter (row-by-row readout) is standard in sCMOS and introduces skew artifacts for fast motion, but is negligible for static specimens.
For rolling shutter, total frame skew = N_rows × line_time. If object displacement during this time is less than one pixel, artifacts are undetectable.
9.Specialized Sensors
Five specialized sensor technologies: EMCCD (single-photon sensitivity), ICCD (nanosecond gating), sCMOS (lowest noise at highest speed), scientific CCD (on-chip binning for spectroscopy), high-speed CMOS (>10⁵ fps).
| Need | Best Technology | Key Advantage |
|---|---|---|
| Single-photon detection | EMCCD | EM gain eliminates read noise |
| Nanosecond gating | ICCD | Photocathode shutter <2 ns |
| Widefield imaging, 10–100 fps | sCMOS (BSI) | Low noise + high speed + large FOV |
| Long-integration spectroscopy | Scientific CCD (LN₂) | On-chip FVB + negligible dark current |
| >1,000 fps | High-speed CMOS | Gigapixel/s throughput, global shutter |
| Photon-number resolving | qCMOS | <0.3 e⁻ read noise, no excess noise |
EMCCD is superior to ICCD in almost every metric except gating speed. If nanosecond time resolution is not required, EMCCD is the better choice.
10.Selection Workflow
Start with the photon budget, then consider temporal requirements, spectral range, and whether the application is spectroscopy (favors CCD) or imaging (favors sCMOS). sCMOS is the default for modern scientific imaging unless a specialized need dictates otherwise.
Beware of dynamic range claims using non-linear full well. Always request EMVA 1288 characterization data for standardized, comparable measurements.
Continue Learning
The Comprehensive Guide includes 7 worked examples, 7 SVG diagrams, 2 data tables, and 10 references.