Machine vision happens in a controlled light box - so why do low-light characteristics matter at all? Because "controlled" rarely means "unlimited". Cycle time caps exposure, depth of field caps aperture, heat caps strobe duty, and the parts cap the wavelength. When the photon budget runs thin, the sensor's ability to use scarce light becomes the difference between an image you can inspect and one you can argue about.
The photon arithmetic
A pixel's signal is the light intensity, times the pixel area, times the exposure, times the sensor's quantum efficiency (QE). Pixel area is the one term that scales with hardware choices. A 3.45 um pixel (the Sony IMX264 in the MV-CS050-10GM) collects (3.45/1.85)^2 = 3.5 times the photons of a 1.85 um pixel (the IMX226 in the MV-CU120-10GM) under identical lighting and exposure. That factor of 3.5 is the difference between comfortable signal and gain-boosted noise in dim conditions.
Quantum efficiency: the quiet multiplier
QE is the fraction of incoming photons the sensor converts to electrons. Modern Sony machine vision sensors run 60-80% at their peak wavelength, with back-illuminated designs pushing higher. QE interacts with wavelength: a sensor strong in the red is the right pick for 660 nm lighting. When comparing cameras, check the QE curve at your light's wavelength, not just the headline peak - a 5% QE difference is invisible, a 20% one is an exposure step.
What low signal actually costs
Photon arrival is statistical: at 100 electrons per pixel, shot noise is about 10%, and no firmware removes it. Underexposed images compensate with analog or digital gain, which amplifies signal and noise together, plus read noise. The visible result: speckled backgrounds where subtle contrast defects - faint scratches, mild discoloration - drown. Defect detection tolerates less noise than the eye does; algorithms threshold on contrast that noise erases.
Practical strategies when light is scarce
Choose big pixels first. If your resolution requirement allows 3.45 um pixels at 5 MP instead of 1.85 um at 12 MP, the 5 MP camera will image the same scene with 3.5x the signal. Resolution you do not need is sensitivity thrown away.
Lengthen exposure before raising gain. Doubling exposure doubles signal with zero noise penalty; doubling gain doubles it with a noise penalty. If motion allows 4 ms instead of 2 ms, take the free lunch.
Match wavelength to QE. Many sensors have an extended response into the near-infrared; IR lighting with an IR-capable sensor often beats visible light on low-contrast marks - and IR light is invisible to operators, a small ergonomic bonus.
Binning, if the interface allows. Combining 2x2 pixels quadruples the signal per "super-pixel" at half the linear resolution - sometimes the right trade for a low-contrast, low-detail task.
Reading this against real cameras
The trio we stock most illustrates the spread: the MV-CS050-10GM (3.45 um pixels, global shutter) is the low-light workhorse; the MV-CS200-10GM (2.4 um, rolling shutter, 1" format) balances resolution with decent area; the MV-CU120-10GM (1.85 um) trades sensitivity for 12 MP in a small format. Same GigE link, same ecosystem - three different photon appetites.
The general rule worth carrying: in machine vision, light is cheaper than pixels. When the spec gets dim, buy photons - bigger pixels, better lighting, longer exposure - before you buy software to denoise what the sensor never captured.