Camera module low light imaging with IR illumination

You need a compact camera that still returns a clean black-and-white image when there is almost no light, and you are free to add your own infrared illumination. The decisions that decide success are fewer than they look: monochrome or color, which infrared wavelength, pixel size and exposure, how the LED is driven, and how frames reach your host. Work through them with arithmetic instead of guesswork and the design settles quickly. This guide supplies the numbers — photons per pixel, field of view, motion blur, duty cycle — ties them into one worked example, and ends with the module decision for each case.

Sensor, module or board-level camera

A bare image sensor converts photons to electrons and little else. Clocking it, powering it, pairing its MIPI CSI-2 lanes and writing its driver is a project of its own. A camera module adds the lens and support parts so a host can pull frames sooner. A board-level camera goes further: a complete industrial camera on a bare PCB, with its own clocking, an M12 or C-mount lens holder, and a USB3 or GigE output that any PC or embedded host already understands.

The host decides the route. Driving a MIPI sensor from a microcontroller without native CSI-2 input means adding a bridge board plus its driver work. A board-level camera replaces that problem with a standard network or USB link, keeps the small footprint, and gives you a documented sensor and mount. The longer checklist — mounting, cabling, sensor format — is in the board level camera selection for embedded vision machines guide.

The MV-CB016-10GM-M-W is that route in one part: a mono sensor with a global shutter behind an M12 mount, streaming over GigE.

Why monochrome wins in the dark

Two filters stand between a color sensor and a dark scene. The Bayer array dyes every pixel red, green or blue, so each pixel passes only its slice of the spectrum — roughly half the photons at best, often less. Above it sits the IR-cut filter, which passes very little beyond about 650-700 nm. Light a scene in the near-infrared and a color camera with that filter in place receives almost nothing, whatever the exposure.

A mono sensor carries neither filter. Every pixel accepts the whole band, so the same lens, scene and exposure return roughly 2x to 4x more signal. In the infrared the margin widens further, because there the color camera starts near zero. Silicon keeps a useful response past the visible edge: mono industrial sensors typically turn a fifth to two-fifths of arriving infrared photons into electrons, and the response rolls off toward 1,000 nm as silicon's absorption weakens.

If you plan to add a bandpass filter matched to the LED, see why a monochrome camera with an optical filter beats color for contrast, and how IR-cut filter options change a color camera's near-infrared reach.

The photon budget, worked end to end

Numbers make the decision concrete. Take a compact build: a mono module with 3 μm pixels in a 1440 × 1080 array behind a 6 mm f/2.0 M12 lens, working distance 500 mm, part speed 100 mm/s, frame rate 60 fps. A 3 μm pixel is typical for this sensor class; substitute your own values and the arithmetic carries over.

The field of view comes first. The sensor is 1440 × 3 μm = 4.32 mm wide, and at a working distance far larger than the focal length:

FOV ≈ 500 mm × 4.32 mm / 6 mm = 360 mm. That scale is 360 mm / 1440 px = 0.25 mm per pixel on the part.

The exposure cap follows from motion. To keep blur inside one pixel, t = 0.25 mm / 100 mm/s = 2.5 ms. At 60 fps the frame period is 1/60 = 16.7 ms, so the 2.5 ms strobe occupies a duty cycle of 2.5 / 16.7 ≈ 15%.

The lens matters as much as the LED. Photon flux through the aperture scales as 1/N², so stopping down from f/2.0 to f/4.0 divides the light by (4/2)² = 4. Work near wide-open and buy range back with LED current instead.

Run the same scene on a color module in the infrared with its IR-cut filter in place and the photon count lands near zero. No exposure recovers that; removing the filter or going mono is the only fix.

When the budget tightens, frame rate buys headroom. LS-BCM-0003 pairs a global shutter with 120 fps and a C-mount on USB3. At that rate the frame period shrinks to 1/120 = 8.3 ms, so strobe and readout must both fit inside it.

850 nm or 940 nm: pick the trade you can live with

Both are standard infrared LED wavelengths, and the choice trades stealth against signal. This section fixes the one number that matters up front: for this class of sensor and emitter, 850 nm is the performance choice.

The shorter wavelength gives the best combination of emitter efficiency and sensor response; its faint deep-red glow is visible in a dark room. Choose it whenever performance outranks concealment. 940 nm is invisible, which is why covert designs use it. The cost is real: the same system delivers up to 40% less usable light, because 940 nm emitters convert less electrical power into radiation and silicon's response keeps falling — typically to around 10% quantum efficiency or less at 940 nm, against 20-40% at the shorter wavelength. Plan on well under half the signal, and recover what you can by opening the lens two stops: f/2.8 to f/1.4 buys back (2.8/1.4)² = 4x the photons.

Infrared bar lights cover both bands, and the light finder filters them by wavelength, length and diffuser.

Pulse the light into the exposure

A dim light that runs continuously spends nearly all its energy between frames, where it does nothing but heat the housing. Spend it inside the shutter-open window instead. Driven at 1 A during the 2.5 ms strobe of the worked example, a light draws 1 A × 2.5 / 16.7 ≈ 0.15 A of average current — the same number of photons on the sensor as continuous 1 A drive, at a fraction of the heat.

LEDs tolerate several times their continuous current when the pulse is short; the pulse-rating curve in the light's data says how hard at your pulse width. Two things turn that headroom into image quality: a controller that fires the light from the camera's strobe output, covered in the LED lighting controller benefits guide, and current control that holds the pulse flat, covered in the LED overdrive and current control guide.

It also needs a global shutter. A rolling-shutter sensor exposes rows one after another, so a frame-timed strobe lights some rows and not others; the shutter-type decision is set out in the global vs rolling shutter sensor guide. Keep the light off whenever the shutter is closed and the dark between frames stays dark.

Getting the frames off the module

A MIPI CSI-2 sensor offers two integration routes. The first is a host with native CSI-2 input — application processors and some system-on-modules. The second is a CSI-2 bridge to a slower bus; it works, but it adds lane pairing, clocking and driver work, and the bridge caps your frame rate. The sizing math is in the MIPI CSI-2 to USB bridge bandwidth guide, and the MIPI CSI to USB UVC bridge checklist lists what a converter must include when the host expects a standard video device.

The third route skips the bridge: a board camera that already speaks USB3 or GigE. The bandwidth check is quick. VGA mono at 8 bits and 60 fps moves 640 × 480 × 8 × 60 ≈ 147 Mbit/s, about 18.4 MB/s — comfortably inside a GigE link, let alone USB3. The MV-CB016-10GM-M-W streams its 1.6 MP frames over GigE with room to spare.

For a USB3 host and optics you choose yourself, MV-CB060-10UM-B is a mono 6 MP bare board on a 1/1.8" sensor with no lens mount — you bring the optics interface.

Where the approach runs out

Three limits deserve respect before you commit.

Ambient infrared. Daylight and incandescent lamps carry a lot of near-IR, and outdoors the sun swamps LEDs. Short exposures help; a bandpass filter matched to the LED, mounted on the lens, helps more — the bandpass filter with red LED light method shows how the pairing blocks ambient light while passing the illumination.

Distance. Irradiance falls with the square of stand-off: double the working distance from 500 mm to 1,000 mm and the light on the part divides by 4. Buy it back with four times the pulse current, within the pulse rating, or four times the exposure, within the blur cap — rarely both exist, which is why range is the first specification to fix.

The noise floor. Below a few dozen photons per pixel the image turns grainy however the light is driven, because read noise takes over. Mono pixels and short, bright pulses push that floor down; cooling pushes it further, but compact boards rarely offer it. And color is gone for good: if any decision in the process needs color, mono plus infrared cannot supply it, and the alternative is white light with a color camera at the sensitivity cost above.

The decision, case by case

• Microcontroller host with no native CSI-2 input: take a mono, global-shutter board-level camera with USB3 or GigE and skip the bridge project.
• Host with native CSI-2 input: a module works; budget the driver and bring-up time before you commit.
• Stealth matters: 940 nm, a lens two stops faster than the 850 nm choice, and half the signal as the planning number.
• Performance matters: 850 nm, global shutter, strobe output driving the light, overdriven pulse inside the exposure.
• Scene wider than a few hundred millimetres: the same arithmetic scales up to area scan cameras with larger sensors and longer working distances.

Next step: filter the camera finder by mono sensor, global shutter and board-level format, then size the light in the lighting configurator. We size the photon budget for your scene if you send the working distance and part speed — the contact page reaches the applications team.