How machine vision cameras are classified, axis by axis

A camera spec sheet reads like code: MV-CS200-10GC, 20 MP, GigE, rolling shutter, colour. Every word on that line is a class on one of eight independent axes, and the axes decide whether a camera can see your defect, keep up with your line and talk to your PC. This guide walks each axis — sensor technology, geometry, scan mode, resolution, output, colour, speed and spectrum — with the rule and the arithmetic that settles it, then closes with a task-by-task decision table. Read it once and you can turn any model string into a plain specification sentence and argue about the single axis that decides the purchase.

The eight axes at a glance

Every machine-vision camera answers the same eight questions, from a 0.3 MP board-level sensor to a 25 MP CoaXPress flagship. The shop categories are the geometry axis split three ways: area scan cameras, line scan cameras and infrared cameras.

A camera is not 'a type'. It holds one position on each axis, the positions are chosen independently, and in practice one axis decides the purchase while the other seven follow from it.

Sensor technology: CMOS took the shelf

CCD sensors move charge to a common output; CMOS pixels convert to voltage where they sit. That is why CMOS reads out faster, absorbs on-chip logic and has replaced CCD in nearly every new industrial design — the IMX and GMAX sensors across the catalog are all CMOS. CCD survives where stacked charge transfer still pays, notably TDI line scan and a few specialist sensors.

The label matters less than the measurement. Current CMOS machine-vision sensors commonly measure 2-4 electrons of read noise and peak quantum efficiency above 60% in EMVA 1288 reports, so pick by data: the EMVA 1288 camera comparison guide explains the columns, and this walkthrough of selecting low-light cameras by EMVA data applies them to a dark scene.

Sensor geometry: area scan versus line scan

An area camera reads every pixel in one exposure: a true snapshot, sized by sensor and lens. A line camera reads one row — or several rows summed as TDI — and builds the image line by line as the part moves past, which removes the width limit: the same 2K sensor scans a 2 m web or a 10 m extrusion. The price is motion discipline. Line scan needs an encoder and steady speed, as the encoder-triggered line scan wiring guide details, and field width follows its own maths, set out in the line scan camera field of view guide. Many area cameras also expose a line-scan readout mode for webs, covered in the area scan line scan mode guide.

Decision: continuous material past a fixed viewpoint — paper, steel, film, grain — goes line scan. Stationary parts, or a camera carried by a robot, go area scan.

A line scan model that carries its own interface family:

Scan mode: global versus rolling shutter

On older datasheets this axis reads interlaced versus progressive; today it reads global versus rolling shutter, and it decides how motion looks. Global shutter exposes all rows at the same instant, so a moving part stays straight. Rolling shutter reads rows in sequence, so successive rows see the part in different places: at 300 mm/s the image skews 0.3 mm for every millisecond of readout span, which destroys a 0.05 mm measurement. Rolling remains fine — often cheaper and more sensitive — when the part is still or when a strobe shorter than the readout span freezes the motion. Global is the default at line speed. The exposure arithmetic for moving webs is worked through in the fps and shutter speed guide for moving webs.

Resolution: pixels per feature, not megapixels

Shop filters split resolution into classes from up to 3 MP through 16-25 MP to over 25 MP, but the class label is not the number that matters. The number that matters is object size per pixel:

Sensor and lens form one system: a 4000-pixel sensor over 64 mm still needs optics that resolve 0.016 mm at the object, which is why the buyer's guide to industrial camera specifications treats resolution as a system number. Set the field width first, then let the calculator below turn pixels per feature into a shortlist.

A camera from the 16-25 MP class:

Output: analog, digital and the interface families

The output axis has two halves: analog versus digital, and within digital, which interface. Analog cameras send composite video down coax to a frame grabber; new builds avoid them, but if you are keeping a legacy line alive, the analog frame grabber integration guide covers the software side.

Digital families separate by bandwidth, cable length and PC hardware:

• GigE Vision — about 1 Gbit/s over up to 100 m of standard networking, and the largest interface group in the catalog; the GigE Vision camera selection guide covers bandwidth, cables and software.
• USB3 Vision — 5 Gbit/s signalling, roughly 3.2 Gbit/s usable, over short cables; sensor-by-sensor trade-offs in the USB3 machine vision camera selection guide.
• 10GigE — ten times GigE for large sensors and long runs.
• CoaXPress — CXP-12 carries 12.5 Gbit/s per link on coax, through a frame grabber.
• Camera Link and XoFLink — line-scan workhorses that need grabber hardware; confirm the pairing with the Camera Link frame grabber compatibility guide.

The deciding sum is pixels x bit depth x frame rate. A 12 MP USB3 model such as the LS-CMR-0166 streams 4000 x 3000 pixels at 30 fps. In 8-bit mono that is 12,000,000 x 8 = 96 Mbit per frame, so 30 fps costs about 2.9 Gbit/s: far past a 1 Gbit/s GigE port, inside USB3's usable 3.2 Gbit/s with thin margin, trivial for 10GigE. At 12-bit depth the same stream rises to 4.3 Gbit/s and USB3 drops out too.

Colour: mono, Bayer colour and beyond

The colour axis trades photons for hue. A mono sensor keeps every photon and interpolates nothing, which is why code reading, gauging and most measurement default to mono. A colour sensor sits behind a 2x2 red-green-green-blue filter array and reconstructs the missing values, so colour detail is partly computed and some light is filtered away — the Quad Bayer 4-cell sensor guide shows what that pattern does to real resolution. Colour is mandatory only when hue carries the information: label verification, food and pharmaceutical sorting, solder inspection. When images show fringes or moire the part does not have, the false colour and moire guide separates sensor artefacts from lighting problems.

Speed class: frame rate, line rate and the blur budget

Area cameras promise frames per second; line cameras promise lines per second, and for moving material only the second tracks honestly. Three sums set the requirement:

Spectrum: UV, visible, NIR, SWIR and LWIR

The last axis is wavelength. Silicon sees roughly 400-1000 nm, so a standard camera is a visible-plus-NIR instrument, and mono sensors reach deeper into the NIR than colour ones because no Bayer filter blocks the deep red. Beyond silicon, three bands do work silicon cannot:

• SWIR, about 900-1700 nm — water absorbs strongly here, and plastics, silicon wafers and hot glass turn transparent or contrasty: moisture grading, fill checks, wafer work.
• UV, below 400 nm — excites fluorescence in adhesives, coatings and surface cracks that vanish under white light.
• LWIR, 8-14 um — images emitted heat, so a thermal camera measures temperature with no external light.

The band-choice reasoning sits in SWIR vs NIR vs visible imaging; when one band is not enough, hyperspectral imaging replaces it with a stack of bands.

Catalog cameras spanning the three non-visible bands:

Choosing by task: the decision table

Fill one position per axis and the shortlist writes itself:

Three checks before ordering:

• Recompute pixels per feature for your own field width; do not inherit a resolution from a previous project.
• Add pixels x bit depth x frame rate and hold the total under the interface's usable bandwidth.
• Match the lens image circle to the sensor format and the mount to the mechanics; a mismatch shows up as dark corners or a working distance that will not fit.

The camera finder filters every axis above directly, and the camera configurator asks the same questions in order; the cameras topic page collects the deeper guides.

When the usual answer is wrong

• More megapixels is better. Only if the lens resolves the extra pixels; below the optics' limit, more sensor pixels add bandwidth and noise, not detail.
• Colour always. Interpolation softens edges and filters cost photons; mono wins wherever hue carries no information.
• CCD is obsolete. TDI line scan and specialist sensors still use charge-transfer designs where they earn their keep.
• Rolling shutter ruins motion. A strobe shorter than the readout span freezes moving parts cleanly on a rolling shutter sensor.

Close the eight axes into one sentence — 'mono, 12 MP class, global shutter, USB3, visible, 20 kHz' — and choosing becomes mechanical. When two axes fight, resolution against bandwidth or spectrum against budget, free expert support with every order covers the trade-off before you commit.