Camera sensor technologies by spectral range: CMOS to microbolometer

A camera sees only the wavelengths its detector physics absorbs, so the spectral range is a sensor decision before it is a software one. Pick the wrong family and the defect you are chasing is genuinely invisible: no lens, light or algorithm recovers it. This guide maps the five detector families used in industrial imaging — silicon CMOS, UV-enhanced backside-illuminated CMOS, InGaAs, cooled MWIR assemblies and uncooled microbolometers — to the bands they cover, the cooling and optics each one forces on you, and a pixel-budget example that shows when a thermal camera cannot resolve a defect at all. It is written for the engineer who has a signal or a defect in mind and needs the simplest detector that still sees it.

Which sensor technology covers each wavelength band?

Each band of the spectrum is served by a different detector: silicon CMOS for 400–1000 nm, UV-enhanced backside-illuminated CMOS below 400 nm, InGaAs for 0.9–2 µm, cooled InSb or XBn assemblies for 3–5 µm, and uncooled microbolometers for 8–14 µm. Two boundaries drive the whole table. Silicon stops absorbing near 1.1 µm, which sets the long edge for CMOS and everything built on it. Below roughly 400 nm, ordinary glass and sensor cover glasses absorb, so UV work needs quartz optics and UV-capable sensors.

X-ray imaging sits outside the table as a scintillator in front of a silicon sensor: the crystal converts X-ray photons to visible light and the CMOS chip reads the result. Everything else in an industrial camera is one of the rows above, and the area scan camera and infrared camera ranges split along exactly these lines.

What can a standard silicon CMOS camera actually see?

A mono silicon CMOS sensor covers 400–1000 nm without modification, so it sees the visible spectrum plus the 850–920 nm band where common IR LEDs emit; colour versions trade that near-IR response for a Bayer mosaic and a debayered RGB image. That single span carries the majority of machine vision: code reading, measurement, presence checks, pick and place. Inside it you choose three things. Mono versus colour, where mono keeps every pixel and the full near-IR response. Interface, GigE for cable length, USB3 for simplicity. And the lens, which must resolve the pixel pitch — matching sensor to lens at the Nyquist limit is that pairing check.

The LS-CMR-0166 is the archetypal VIS/VNIR workhorse: 12 MP of silicon in 1/1.7" format, 4000 × 3000 pixels read out at 30 fps over USB3. For most plants, this row is the whole answer, and the families below exist only for the signals silicon cannot absorb.

When does UV imaging need more than a standard sensor?

Standard glass and standard silicon both fade below about 400 nm, so UVA work at 315–400 nm runs on BSI mono sensors, while UVB at 280–315 nm and UVC below 280 nm need UV-enhanced sensors engineered for short-wave quantum efficiency. Backside illumination is the enabling change: flipping the sensor puts the photodiode in front of the transistor stack, so UV photons are absorbed before the metal layers eat them. On a BSI mono camera, UVA fluorescence is now an ordinary industrial build. Below 315 nm, quantum efficiency falls away in a conventional silicon stack, and UV-enhanced sensor designs are the practical route. The optics change with the band: glass absorbs short-wave UV, so lenses go quartz or UV-corrected — what changes in a lens outside visible light covers the materials.

Why does SWIR need InGaAs instead of silicon?

Silicon stops absorbing near 1.1 µm, so imaging 0.9–2 µm needs indium gallium arsenide, a hybrid detector in which an InGaAs photodiode layer is bump-bonded to a silicon readout circuit similar to a global-shutter CMOS sensor. Off-the-shelf industrial InGaAs cameras now reach around 5 MP, but three costs come with the band. Pixel pitch runs 10–20 µm against 2–5 µm for modern silicon, so the same format carries far fewer pixels. Dark current is much higher, so the detector is thermoelectrically cooled tens of degrees below ambient, and long-exposure scientific builds go far deeper. And InGaAs detectors cost several times their silicon equivalents.

Both cards are GigE industrial SWIR cameras: the LS-IRC-0029 with 0.3 MP on 15 µm × 15 µm pixels, the fanless MV-CI013-GS-NN at 1.3 MP on the IMX990 sensor. Choose InGaAs when the contrast only exists up there — moisture absorption, wafers under opaque films, polymer sorting past 1.1 µm — starting from which band to inspect with: SWIR, NIR or visible. Long-exposure builds add drift questions, covered in uncooled SWIR camera stability.

Why is MWIR a cooled detector assembly, not a camera sensor?

At 3–5 µm the dark current of suitable semiconductors is so high that the detector must be chilled: InSb runs near 77 K (−196 °C) behind an integral cryocooler, XBn-InAsSb works around 150 K, about −123 °C, so you buy a sealed assembly, not a board-level sensor. The 3.6–4.2 µm band is where XBn barrier detectors earn their complexity, and InSb stays the reference for high-speed thermal events. What justifies the cooler: gas imaging, since hydrocarbon bonds absorb strongly near 3.3 µm and a leak reads as a hot plume; black plastics, which carbon-black loading makes indistinguishable in NIR sorting but separable in MWIR; and active thermography on metals, where heat diffuses fast enough that only a high-frame-rate cooled imager catches the crack.

MWIR is bought as an engineered detector unit with its own cryocooler and control electronics, so the models are quoted items rather than shelf cameras — plan the lead time and the mounting heat load into the project.

How does an uncooled microbolometer measure heat at 8–14 µm?

A microbolometer is a thermal detector: LWIR radiation at 8–14 µm warms a membrane whose resistance changes, giving 30–40 mK NETD at 30–60 Hz without any cryogenic cooling, which is why every industrial thermography camera uses one. Because the readout is thermal, not photonic, the trade-offs are temporal and spatial: response is relatively slow, arrays run QVGA to VGA with megapixel parts appearing, and the germanium lens is calibrated to the sensor as one unit, so optics are not swappable afterwards. The common temperature span for industrial radiometry is about -10 to 140 °C, and 9 Hz variants exist chiefly to stay under export-control thresholds.

The LS-IRC-0044 is the class in one card: 640 × 512 pixels on an 8–14 µm band, streamed at 50 Hz over GigE through an 88.5° × 73.2° lens. For monitoring work, NETD and lens choice for process monitoring sets the deciding numbers; for measurement, raw thermal counts to temperature is the calibration step after the optics.

Can a thermal camera resolve your defect?

Often no: budget three pixels across the smallest feature, divide field of view by pixel scale, and a 640 × 512 microbolometer on a 100 mm field reaches only 0.16 mm/px, so any feature under about 0.5 mm is gone regardless of its 30–40 mK sensitivity. The arithmetic is two lines: pixel scale equals field of view divided by pixels across the sensor, and a feature needs about three pixels to be detected reliably.

Worked example: a 0.25 mm crack on a 100 mm wide machined face.

• Required scale: 0.25 mm ÷ 3 = 0.083 mm/px.
• Required array: 100 mm ÷ 0.083 mm/px ≈ 1200 px across.
• Silicon at 4000 px: 100 mm ÷ 4000 = 0.025 mm/px, so the crack spans 0.25 ÷ 0.025 = 10 px. Detected and measurable.
• Microbolometer at 640 px: 100 mm ÷ 640 ≈ 0.16 mm/px, so the crack spans 0.25 ÷ 0.16 ≈ 1.6 px. Not detectable, whatever the NETD table says.

The arithmetic lands on InGaAs too: at 15 µm pitch an InGaAs sensor carries about a third of the pixels per line of a 5 µm silicon part in the same format, so the SWIR field of view shrinks or the magnification drops. Run your own numbers in the camera resolution calculator.

Which sensor technology should you choose?

Choose the least exotic detector that still sees the defect: silicon for 400–1000 nm, UV-enhanced BSI below 400 nm, InGaAs from 0.9 µm, cooled assemblies only when 3–5 µm absorption or fast thermal events demand it, and microbolometers whenever the signal is surface temperature. Routing by case:

• Code, geometry, colour, surface: silicon CMOS, mono unless you truly need colour — the camera finder filters by sensor, resolution and interface, and the camera step of the configurator builds the chain around your pick.
• UVA fluorescence, 315–400 nm: BSI mono camera with quartz optics.
• UVB and UVC below 315 nm: UV-enhanced sensor, quartz optics, interlocked illuminator.
• Contrast only above 1.1 µm, moisture, wafers, polymer sorting: InGaAs SWIR.
• Gas, black plastic, fast thermal events: cooled MWIR assembly.
• Surface temperature from about -10 to 140 °C, hot spots, delamination by cooling behaviour: microbolometer LWIR.
• Material identity rather than presence: hyperspectral cameras splitting the silicon band itself — the LS-HSC-0004 resolves 400–1000 nm into 1200 bands over USB3 — with how hyperspectral imaging works as the method check.

Two checks kill more spectral builds than sensor choice does. Lens transmission: glass stops near 1.1 µm on the long side and below about 350 nm on the short side, so quartz or germanium optics are part of the order, not an accessory. Cooling duty: a thermoelectrically cooled SWIR camera or a cryogenically cooled MWIR unit must reject its own heat, which decides the mounting before the first image. If two rows still look plausible, re-run the pixel budget above, or send the part and the defect to expert support — we will match band, array and interface in one pass. The cameras topic page collects the deeper guides on each family.