CCD vs CMOS sensor differences: what a machine vision buyer checks

Every machine vision camera you can buy new today converts light on a CMOS sensor; CCD imagers, the default for three decades, have left mainstream builds. The label still matters when you read a legacy drawing, swap a failed camera or chase a uniformity argument, and the physics behind it explains every number on a modern data sheet. This guide separates what the CCD-to-CMOS transition really changed from what it never decided, works the numbers that do decide image quality — quantum efficiency, pixel pitch, full well and read noise — with arithmetic you can check, and closes with the five checks to run before you order.

What is the real difference between CCD and CMOS sensors?

A CCD shifts every pixel's charge packet to a single output and converts it once, which gives high pixel-to-pixel uniformity at a readout speed the charge transfer itself sets; a CMOS pixel converts its own charge to voltage at the site, and rows read out in parallel.

A CCD is a bucket brigade: clock pulses walk each pixel's charge along shift registers to one output node, where a single amplifier and one ADC turn it into a voltage. Every electron sees the same conversion, so uniformity was excellent, and the same transfer set the speed limit: packets move at clock speed and no faster. Finite well depth meant a bright spot could spill into neighbouring pixels — the smear called blooming. How that shared ADC's bit depth reaches your image is its own story, told in why 10-bit sensors ship 8-bit images.

A CMOS pixel converts where the light lands: a photodiode plus a few transistors reset, amplify and select each site, and column ADCs digitise rows in parallel. Speed, low power and a smaller camera head follow. The old costs — pixel-to-pixel variation, fixed-pattern noise, lower fill factor — were real, and modern fabrication, per-column calibration and microlenses have closed most of that gap.

Why did CMOS replace CCD in machine vision cameras?

Three practical wins decided it: pipelined readout that exposes frame n+1 while frame n streams, on-chip digitisation that shrinks the camera head, and power low enough for compact, cooler housings — a 12 MP head today delivers 4000 × 3000 at 30 fps over one USB3 cable.

Blooming finishes the argument. A saturated CMOS pixel clips locally instead of spilling charge down a column, so hot weld seams and glowing filaments image without smear, and high-dynamic-range modes arrived on the same chip. Catalogue reality backs it: filter area scan cameras by shutter and 904 models return as global-shutter, across interfaces from GigE to CoaXPress. The mainstream head in one card: the LS-CMR-0166, 12 MP, 4000 × 3000 at 30 fps over USB3, on a 1/1.7" rolling-shutter IMX226.

Which sensor numbers matter more than the CCD or CMOS label?

Four numbers decide what your camera sees: quantum efficiency at your lighting wavelength, pixel pitch against your lens resolution, full well against your brightest surface, and read noise against your darkest one. Every one is printed in the sensor data, and none follows from the CCD or CMOS tag.

They live in the sensor's EMVA 1288 characterisation — QE plotted against wavelength, temporal dark noise, saturation capacity, dynamic range in dB. Compare two cameras on those rows and the family label adds nothing; the curve at your wavelength decides. Reading EMVA 1288 data shows where each figure comes from.

Dynamic range is the column that closed: CCDs owned it in the 2000s, while current machine-vision CMOS prints around 60-70 dB as a matter of course. The decision has moved onto the data sheet, so read it there. A current 20 MP mono part, IMX183, on a 1" format over USB3.0:

How does quantum efficiency set your wavelength choice?

Quantum efficiency is the share of photons that become electrons — ten photons in, four electrons out is 40% QE — and silicon peaks near 30-60% in the visible, falls past 900 nm and stops near 1100 nm, where the bandgap ends the response.

The physics is the bandgap: a photon needs more energy than the gap to free an electron, so the cutoff wavelength follows λ ≈ 1240 / Eg, with Eg in electron-volts. Silicon's 1.1 eV gap walls the response near 1100 nm; InGaAs' ~0.75 eV gap reaches about 1650 nm, which is why SWIR cameras build on InGaAs photodiode arrays with CMOS-style readout. Past the infrared, imaging turns thermal: an uncooled microbolometer array reads 7-14 μm, each pixel a resistor whose resistance shifts with absorbed heat — the thermal camera NETD guide covers what that means for lens choice.

Two buying consequences. First, read QE at your LED's peak wavelength, not the headline peak: silicon that returns 60% at 550 nm can fall under 10% at 940 nm, and the same 850 nm drive gives visibly more signal — choose 940 nm only when the glow must be invisible. The near-infrared wavelength guide works this trade for real surfaces. Second, color sensors need the near-infrared removed: an IR-cut filter keeps NIR out of the red channel, and choosing that filter glass is a real design step, not a checkbox.

When the material sets the band, buy the sensor for the band: the LS-HSC-0004 hyperspectral camera spans 400-1000 nm with 1200 bands over USB3, silicon's whole usable window, while the LS-LCM-0038 covers 900-1700 nm at a 12.5 μm pitch and 29.4 kHz — the band silicon cannot reach. Both live in hyperspectral cameras and line scan cameras.

How does pixel size trade resolution against light?

Pixel pitch sets two limits at once: the finest detail the sensor can sample, roughly 500 divided by pitch in μm and read in line pairs per millimetre, and the charge each site holds, which grows with pitch squared. A 7 μm pixel resolves about 71 lp/mm; a 5 μm pixel about 100 lp/mm.

The squared term is the light side: bigger wells saturate later and start from a higher signal, which is why low-light and SWIR parts run fat pixels while small-pixel heads lean on the lens and the light. The detail term bites the lens: a sensor cannot sample what the lens never resolved, and a 2.4 μm pitch asking 208 lp/mm will photograph every flaw in a mediocre optic — small-pixel camera selection walks that negotiation, and crosstalk plus microlens angular tolerance set the practical ceiling.

Line scan shows the span in one row of data sheets, because pitch is printed there: the MV-CL166-91F2C reads 16K lines at a 5 μm pitch over XoFLink and needs an 82 mm image circle; the LS-LCM-0016 runs 2K at 7 μm and 58 kHz; the LS-LCM-0038 covers 900-1700 nm at 12.5 μm:

Values as printed in each product's data sheet; open a part for its drawing and full specifications.

How many pixels does your defect actually need?

Work backwards from the part, not the sensor: a 110 mm field inspecting a 0.1 mm defect at three pixels across needs 110 ÷ (0.1 ÷ 3) ≈ 3300 pixels, which a 1600-pixel sensor fails and a 4000-pixel sensor passes.

The arithmetic, end to end. A machined part 100 mm wide takes a 10% framing margin, so the field of view is 110 mm. The smallest defect that matters is 0.1 mm, and three pixels across it is the floor for a dependable classifier, so the scale must reach 0.1 ÷ 3 ≈ 0.033 mm per pixel.

Pixels needed: 110 ÷ 0.033 ≈ 3300. A 2 MP sensor at 1600 pixels across covers 110 mm at 110 ÷ 1600 ≈ 0.069 mm per pixel, so the defect lands on 0.1 ÷ 0.069 ≈ 1.5 pixels — one pixel of edge, fail. The LS-CMR-0166 from above spans 4000 pixels across. At that span, 110 ÷ 4000 ≈ 0.0275 mm per pixel, the defect sits on 0.1 ÷ 0.0275 ≈ 3.6 pixels, and you hold 700 pixels of horizontal margin for gain, vibration and part wander.

Carry your own part through the resolution calculator, then check the lens still resolves that scale:

Does CCD or CMOS decide whether you get a global shutter?

No — shutter type is a readout architecture, not a sensor family: rolling-shutter CMOS exposes rows in sequence, global-shutter CMOS adds transistors that sample every pixel at once, and every fast camera you buy today is global-shutter CMOS. Match the shutter to motion, then set exposure from speed and pitch.

A rolling shutter exposes each row as it clears for readout, so a part moving mid-scan comes out skewed or sheared; a global shutter stores every pixel in the same instant and reads afterward, paying extra pixel transistors and somewhat less well depth. CCDs were global by construction — readout happened after exposure — but their transfer speed could not feed modern frame rates. Geometry is the risk, and global versus rolling shutter shows the artifacts.

Exposure comes from speed and pitch. A belt at 500 mm/s under a 3.45 μm-pitch sensor blurs one pixel at exposure t = 3.45 μm ÷ 500 mm/s ≈ 7 μs — short, but a strobed pulse covers it easily, and exposure time versus shutter speed sets out the balance.

When you need simultaneity and speed together, global-shutter CMOS is the shelf answer: the MV-CH120-11UC pairs a 12 MP global-shutter IMX304 with a high-speed USB3 readout on a 1.1" format.

Where do CCDs still appear in machine vision?

Only in legacy and niche builds: analogue cameras, some Camera Link frames, and a few scientific or low-light designs where extreme uniformity justifies slow readout. New area-scan and line-scan catalogues list CMOS exclusively, so a failed CCD camera is now a replacement decision, not a repair.

The replacement is a matching exercise, not an upgrade: hold the horizontal and vertical resolution, the optical format so the lens circle still covers the sensor, the pixel size so magnification and pixel calibration stay valid, the mount, and the interface the frame grabber speaks. A step-by-step legacy CCD replacement guide walks that swap for robot-vision retrofits.

The one CCD-born idea that mainstreamed is time delay integration. TDI adds the charge from a moving line image coherently across successive stages, so a fast web collects the photons of many lines at the read noise of one — sensitivity without slowing the belt. Modern CMOS line scan implements TDI on-chip, and the scale arithmetic is unchanged.

Take a web at 1 m/s inspected at 0.05 mm per line: the camera must deliver 1000 ÷ 0.05 = 20,000 lines/s = 20 kHz. The LS-LCM-0016 is a 2K GigE line scan camera rated at 58 kHz with 7 μm pixels and TDI. The web's 20 kHz sits at about a third of that line rate, leaving margin for speed drift; setting the line rate and encoder ticks per pixel finish the set-up.

Which sensor checks do you run before ordering?

Run five checks in this order: QE curve against your light's peak wavelength, pixel pitch against your lens MTF, optical format against the lens image circle, shutter type against part motion, and frame rate against link bandwidth. Each has failed more systems than the CCD-versus-CMOS choice ever did.

• QE at your wavelength. Overlay the sensor QE curve with your LED's peak; the 850 nm versus 940 nm gap is real signal, and some heads drop under 10% at 940 nm.
• Pitch against the lens. The lens must resolve near 500 ÷ pitch in lp/mm at your field; a 2.4 μm pitch asks 208 lp/mm, beyond many C-mount lenses at working distance, so read the MTF curve before the megapixel count.
• Format against the image circle. The sensor diagonal must fit the lens circle or the corners fall off — the tunnel effect; the inch names are legacy pickup-tube conventions, not dimensions, so read the printed diagonal. 25 MP-class heads moved to F-mount and M58 mounts for exactly this reason.
• Shutter against motion. Global when the part moves during exposure, rolling when nothing does; blur at or under one pixel means exposure ≤ pitch ÷ speed.
• Frame rate against bandwidth. Pixels × bytes × fps must fit the link: 4000 × 3000 × 1 byte × 30 fps ≈ 360 MB/s, past GigE's roughly 110 MB/s after overhead — which is why that class of head ships on USB3, and GigE Vision camera selection covers the cable side.

Run the five and the CCD-versus-CMOS question answers itself, because you never asked it: the numbers chose the camera. Shortlist in the camera finder, compare EMVA 1288 rows the way the buyer's guide to industrial camera specifications sets out, then configure the camera once the sensor checks pass. The rest of the sensor-side guides sit on the cameras topic page.