Image Sensor Selection for Machine Vision: Resolution, Shutter Type, and Lens Pairing by Application

Global shutter vs rolling shutter for machine vision

Global shutter exposes every pixel at the same instant; rolling shutter exposes rows sequentially over a readout period lasting microseconds to milliseconds. For a stationary scene, the difference is invisible. For anything that moves relative to the camera (the object, the camera, or both), rolling shutter introduces geometric distortion because the scene changed between when the first row and the last row were sampled.

Rolling shutter failure modes

A vertically oriented object moving horizontally shears into a parallelogram: a rectangular part on a conveyor no longer measures as rectangular. On a sensor with a 10ms readout, a part moving at 500mm/s moves roughly 5mm in object space during readout; the resulting image-plane smear or skew is that displacement times the system magnification. Camera-mounted vibration produces a wavy, jello-like wobble as different rows sample slightly different angular positions. This corrupts SLAM and visual odometry, which depend on consistent pixel positions frame to frame.

Strobed illumination compounds the problem. If a strobe fires faster than the full readout period, only the rows being read at that instant receive the flash, producing a bright band across some rows and dark rows above and below it. Synchronizing a short strobe pulse with rolling shutter defeats the purpose of the short pulse: the strobe duration would need to span the entire readout to expose evenly.

When global shutter is required

Conveyor inspection, robotics and pick-and-place, strobed illumination, precision dimensional measurement, and any system where the camera itself moves during operation all require global shutter. Even a slow 200mm/s conveyor belt moves 0.2mm during a 1ms readout. For a 10µm feature, that is a real positional error at typical machine vision magnifications, not a rounding error. Barcode decoders depend on consistent bar spacing; rolling shutter skew relative to the direction of travel can distort bar spacing and cause decode failures on fast-moving lines.

When rolling shutter is good enough

Rolling shutter is not a defective technology. It is the right choice when the scene does not move during the readout period. Document capture, static QR and label reading, and kiosk scanning involve no object motion during exposure. Web inspection of a continuously moving material can still use rolling shutter if the camera is triggered with an exposure short enough that the expected skew from web speed times readout time stays within your measurement tolerance; calculate it rather than assume it. Cost-sensitive embedded modules for retail analytics or presence detection often use rolling shutter deliberately, because the scenes are static or near-static and the artifact is below the threshold the application needs to distinguish.

A sharper or lower-distortion lens will not reduce rolling shutter skew or flash banding, because the distortion is a temporal readout artifact, not an optical aberration. Fixing it requires a global shutter sensor or a guarantee that the scene stays still for the full readout period. Shutter type also correlates with lens mount in practice: global shutter cameras in higher-accuracy industrial systems commonly pair with C-mount lenses, which give an adjustable iris ring for depth-of-field control. That control is practical because illumination in these systems is usually programmatically controlled. Rolling shutter sensors dominate compact embedded modules paired with M12 lenses for size and weight.

See the M12 lens collection and the M12 vs C-mount vs CS-mount guide for the full mount tradeoff.

NIR sensitivity and illumination strategy

Sensor spectral response should follow the illumination strategy, not the other way around. Standard silicon CMOS sensors retain meaningful quantum efficiency into the near-infrared, typically out to 1000–1100nm, but most machine vision cameras ship with an IR-cut filter installed to preserve visible-light color accuracy. If your system illuminates with 850nm or 940nm LEDs (common for covert lighting, low-visible-light environments, or combined day and night operation), remove the IR-cut filter or specify a NIR-optimized variant, and confirm the sensor's QE curve at your chosen wavelength rather than assuming uniform NIR sensitivity across parts.

850nm sensors retain higher QE than 940nm variants on most silicon, at the cost of a faint visible red glow from the illuminator that is sometimes undesirable in public-facing installations. 940nm is effectively invisible to the human eye but loses roughly half the QE of 850nm on typical CMOS sensors, which means either brighter illumination or a faster lens aperture to compensate. Match sensor, filter, and illuminator wavelength as one decision, not three independent ones: a NIR-sensitive sensor behind a standard visible bandpass filter gains nothing from the illuminator.

Sensor format and lens coverage

Sensor format is the physical size of the imaging area. The diagonal measurement sets how large a lens image circle you need. If the image circle is smaller than the sensor diagonal, the corners receive no light, producing vignetting. Larger sensor formats capture a wider field of view at a given focal length, or let you use a longer focal length to hold the same field of view with a shallower depth of field. See CMOS sensor size for the full format-to-dimension reference and sensor size and lens compatibility for the vignetting math.

A lens rated for a 2/3" sensor produces visible corner vignetting on a 1.1" sensor at every aperture. Stopping down does not fix a coverage mismatch, because the problem is geometric, not a depth-of-field effect. Always match or exceed the sensor format with the lens specification; oversizing the lens format works fine but may add unneeded cost.

Dynamic range, quantum efficiency, and EMVA1288

Dynamic range

Dynamic range measures how many stops of light intensity a sensor captures without clipping highlights or losing shadow detail. It matters most in scenes with simultaneous bright and dark regions: shiny metal and black rubber in the same frame, or an outdoor inspection under direct sunlight with shadowed areas. Many Pregius sensors deliver 60–70 dB of dynamic range. HDR modes exist on some sensors but add complexity to illumination control and image processing; standard dynamic range is typically sufficient unless you have a documented clipping problem.

Quantum efficiency

Quantum efficiency (QE) is the fraction of incident photons that generate an electron in the photodiode. A sensor with 80% QE generates twice the signal of a sensor with 40% QE under identical illumination. Backside-illuminated sensors (Sony Pregius S) typically show higher QE than front-side illuminated sensors at the same pixel pitch, because the backside construction removes the metal wiring layers that partially block the photodiode in front-side designs. Higher QE reduces required illumination power and improves signal-to-noise ratio in low-light applications.

EMVA1288

EMVA1288 is the European Machine Vision Association standard for measuring and reporting sensor performance: quantum efficiency, dark current, temporal noise, fixed-pattern noise, dynamic range, and sensitivity, all under a defined methodology. Vendor headline specs are not directly comparable across manufacturers because measurement conditions vary. When a sensor decision is high-stakes, request EMVA1288 characterization data rather than comparing marketing datasheets.