Cameras & sensorsExplainer

NIR band from an RGB camera: what you get without the IR-cut filter

Removing the IR-cut filter from an RGB camera does not free a NIR band. What the Bayer sensor really captures, and the rig that gives a true NDVI.

LumaSpec Vision12 min read Share on LinkedIn

On this page 10 sections
  1. 01What the sensor measures once the filter is gone
  2. 02Why no code can split an NIR band out of Bayer data
  3. 03Route 1: one camera, two filters, two captures
  4. 04Route 2: a beam splitter captures both bands at once
  5. 05Route 3: two cameras side by side, and the parallax you must budget
  6. 06The exposure budget in the NIR band, worked
  7. 07When two bands are not enough
  8. 08Which route fits your case
  9. 09Checks before you order
  10. 10Questions

You removed the IR-cut filter from an RGB camera, or you are about to, because you want the infrared band the way your code already returns red, green and blue. The direct answer: that band does not exist inside a Bayer camera. The color filter array sits on the pixels and separates visible light only, so with the filter gone, every NIR photon lands in all three channels at once. Below are the physics behind that, the reason the single-camera NDVI shortcut collapses, and three routes that produce a genuine NIR band — sequential filter swaps, a beam-splitter dual camera, or a hyperspectral camera — with the arithmetic to pick one for your scene speed and depth range. If you are building a plant-health station or any band-ratio inspection, every decision here starts from those numbers.

What the sensor measures once the filter is gone

A silicon imaging sensor converts photons until their energy drops below the silicon bandgap of 1.12 eV; past that point the chip is blind. The cutoff follows directly from the bandgap:

λ_max = 1240 eV·nm / 1.12 eV ≈ 1107 nm

Every stock camera fits an IR-cut filter, also called a hot mirror, that blocks roughly 700-1100 nm so recorded color matches human vision. Remove it and that whole range reaches the silicon again. The response is not flat: typical silicon sensors hold about 10-30% of their visible-band response at 850 nm and fade toward zero by 1100 nm, which is why converted cameras shoot dark NIR frames. The exact curve differs per sensor, so read the vendor's spectral response curve before you budget exposure.

The Bayer dyes stay on the chip. Each dye passes some NIR on top of its visible color, so in that band all three channels read similar gray levels, usually with red highest. You get one monochrome-looking image with a color bias, not a set of separable bands.

Choose the wavelength before you buy filters. 850 nm is the default: response is still useful, long-pass filters and LED floods at that wavelength are commodity items, and leaf reflectance is already near its plateau. Going to 940 nm costs extra exposure for little index gain; pick it only when the illumination must stay invisible. Other band choices are compared in near infrared wavelength selection for water and silicon inspection, and the two filter types are measured head to head in IR pass vs IR cut lens sensitivity. If the goal is occasional NIR imaging rather than a production station, the switchable filter options are laid out in color camera near infrared imaging: IR-cut filter options.

On many sensors the IR-cut is part of the cover-glass stack, so removal changes back-focus and dust sealing as well as transmission; what to expect from a bare sensor is covered in cameras without sensor cover glass.

Why no code can split an NIR band out of Bayer data

Channel-splitting code works on data the optics already mixed. The three images you get back are near copies of one another, not three bands, because demosaicing interpolated every pixel across its neighbors before your code saw a value.

NDVI — the Normalized Difference Vegetation Index — is where this fails visibly. The index contrasts NIR against red:

NDVI = (NIR - Red) / (NIR + Red)

Healthy leaves reflect weakly in red and strongly beyond 700 nm; reflectance typically climbs from under 10% in red to 40-50% just past 750 nm, and the index reads exactly that contrast. In a converted camera, the red channel is itself NIR leaking through the red dye, so the numerator trends to zero and the output goes flat regardless of plant health. For orientation, remote-sensing rule values sit near 0.1 or below for soil and water, 0.2-0.4 for sparse vegetation and 0.6-0.9 for dense canopy.

Sensors with a fourth NIR pixel class (RGBN) do compute NDVI from one frame, but they are rare in machine vision. If you are weighing a narrow NIR bandpass on a color sensor instead, know that narrow filters trade band purity for bit depth; the arithmetic is worked through in narrow bandpass filters on an RGB sensor.

Route 1: one camera, two filters, two captures

The simplest rig swaps filters between frames: an 850 nm long-pass for the NIR frame, then the stock arrangement or a clear window for the visible frame. A motorized filter wheel in front of the lens automates the swap and holds the geometry fixed.

The scene must hold still between captures: stop the conveyor, or the two bands will not align, and leaf motion in wind will bias the index. Exposure is the second constraint. With only 10-30% of the visible response left at that wavelength, plan on 5-10× the visible exposure at the same f-number. Lock exposure, gain and iris manually so each band cannot rescale itself; the automatic iris, exposure and gain control loop shows which settings to pin.

Route 2: a beam splitter captures both bands at once

A 50/50 beam-splitter cube, or a matched prism block, sits behind one lens and feeds two synchronized cameras: the color camera keeps its IR-cut for the visible band, and the mono camera runs a matched long-pass. Both sensors share one optical path, so the bands register within a few pixels at any depth. This is the standard answer for moving belts and per-part indices.

Two practical numbers. A 12.7 mm cube shifts focus by t · (1 - 1/n) ≈ 4.2 mm at n = 1.5, and each 3 mm of filter glass adds about 1 mm of path, all inside a C-mount flange distance of 17.526 mm; it fits, but budget it. Second, both cameras must trigger from one source, because two rolling shutters scan at different moments unless they share one exposure event.

A matched sensor pair keeps gain and response alike across bands. The Hikrobot MV-CS200-10GC and MV-CS200-10UM pair shares a 20 MP sensor in 1" format on C-mount with a rolling shutter, color over GigE and mono over USB3:

Specification MV-CS200-10UM MV-CS200-10GC
Max. frame rate 19.2 fps @ 5472 × 3648 Mono 8 5.9 fps @ 5472 × 3648
Exposure time 44 μs to 0.83 sec 46 μs to 2.5 sec
Exposure mode Off/Once/Continuous exposure mode, supports Global Reset and Rolling Off/Once/Continuous exposure auto, supports Timed and Trigger Width exposure mode, and Rolling and Global Reset sensor shutter mode
Pixel format Mono 8/10/10Packed/12/12Packed Mono 8/10/12, Bayer GB 8/10/10p/12/12p, YUV422Packed, YUV422_YUYV_Packed, RGB 8, BGR 8
Decimation Supports 1 × 1, 2 × 2, 4 × 4 Supports 1 × 1, 2 × 2
Reverse image Supports horizontal and vertical reverse image output Supports horizontal and vertical reverse image output
Data interface USB3.0, compatible with USB2.0 Gigabit Ethernet, compatible with Fast Ethernet
Digital I/O 6-pin P7 connector provides power and I/O, including opto-isolated input × 1 (Line 0), opto-isolated output × 1(Line 1), bi-directional non-isolated I/O × 1 (Line 2) 6-pin Hirose connector provides power and I/O, including opto-isolated input × 1 (Line 0), opto-isolated output × 1 (Line 1), bi-directional non-isolated I/O × 1 (Line 2).
Power supply 9 VDC to 24 VDC, supports USB3.0 power supply 9 VDC to 24 VDC, supports PoE
Power consumption Typ. 2.3 W @ 5 VDC (USB3.0 provides power supply) Typ. 2.5 W @ 12 VDC
Dimension 29 mm × 29 mm × 30 mm (1.1" × 1.1" × 1.2") 29 mm × 29 mm × 42 mm (1.1" × 1.1" × 1.2")
Weight Approx. 80 g (0.18 lb.) Approx. 100 g (0.22 lb.)
Price $759.00 $739.00

All 2 share: Sensor type CMOS, rolling shutter; Sensor model Sony® IMX183; Pixel size 2.4 μm × 2.4 μm; Sensor size 1"; Resolution 5472 × 3648; Dynamic range 65.5 dB; SNR 41.5 dB; Gain 0 dB to 24 dB. Values as printed in each product's data sheet; open a part for its drawing and full specifications.

Route 3: two cameras side by side, and the parallax you must budget

Two same-family cameras with the same lens model on one baseplate is the quickest layout to build. Undistort each camera, then fit a homography that maps the NIR frame onto the visible frame over the working plane; the procedure is walked through in multispectral camera band registration. Calibration removes the fixed offset; what survives is the depth-dependent shift:

e = f · B · ΔZ / Z²

where f is focal length, B the baseline between lens centers, Z the working distance and ΔZ the depth spread of the target.

Worked example: two 5 MP 2/3-inch sensors with 3.45 µm pixels, 12 mm lenses, B = 100 mm, seedlings at Z = 400 mm with ±30 mm of leaf-height spread. Sensor-plane shift: e = 12 × 100 × 30 / 400² = 36,000 / 160,000 = 0.225 mm, which is 0.225 / 0.00345 ≈ 65 pixels. Per-leaf NDVI is noise at that offset. Raise the camera and narrow the base, Z = 800 mm and B = 50 mm: e = 12 × 50 × 30 / 800² = 18,000 / 640,000 = 0.028 mm, about 8 pixels, workable for a block-averaged canopy index. Flatten the scene instead, ΔZ = ±2 mm at Z = 400 mm and B = 100 mm: e = 2,400 / 160,000 = 0.015 mm, about 4 pixels, and side-by-side wins on simplicity.

Decision rule from those numbers: depth spread under about 5 mm at working distance means side-by-side is safe; tens of millimeters means a beam splitter; per-leaf indexing means only shared-optics routes qualify.

On a moving belt, global shutter on both cameras freezes the same instant. The MV-CU020-90GC is a 2.5 MP color camera with a 1/1.7" sensor, C-mount, global shutter and GigE output:

Area scan camerasHikrobot MV-CU020-90GC 2.5 MP Area Scan Camera, GigE, GMAX4002, ColorMV-CU020-90GC
Sensor format
1/1.7"
Lens mount
C-mount
Interface
GigE
Resolution
Up to 3 MP
$309.00View partBuild a system with it

The exposure budget in the NIR band, worked

Take a visible frame you are happy with at 2 ms and f/4. Fit the long-pass with about 90% transmission; residual sensor response at 850 nm is about 20% of the red band. The NIR exposure becomes t = 2 ms / (0.9 × 0.2) ≈ 11 ms. On a belt at 100 mm/s that is 1.1 mm of smear, roughly 320 pixels at 3.45 µm, which is unusable. Opening the lens to f/1.8 buys (4 / 1.8)² ≈ 4.9× more light and lands near 2.2 ms: smear drops to about 65 pixels at 100 mm/s, still too much, but about 7 pixels at 10 mm/s, fine for a canopy-level index.

The decisions follow. At 10 mm/s or on stopped parts, aperture and exposure carry the job. At 100 mm/s or faster, strobe an NIR flood so a 2 ms exposure stands; you buy light, not time. A beam-splitter rig needs the same short exposures because both bands must freeze the same instant. Whichever route you take, refocus with the final filter fitted: NIR focuses at a different plane, which is why visible and SWIR bands drift off one focus, and why IR-corrected machine vision lenses exist for two-band designs.

When two bands are not enough

Once the question grows past NDVI into moisture, disease, ripeness or material identity, Bayer leakage cannot carry it; you need dedicated bands. A hyperspectral camera samples the spectrum directly: the LS-HSC-0004 covers 400-1000 nm in 1200 bands over USB3 on a C-mount with a 5 MP sensor.

Hyperspectral cameras400-1000 nm area scan hyperspectral camera, 1200 bands, USB3 (LS-HSC-0004)LS-HSC-0004
Lens mount
C-mount
Interface
USB3
Resolution
5 MP
$15,900.00 · quotedView and quote

How this camera class works, and where it beats RGB cameras and filter wheels, is set out in hyperspectral imaging: how it works and when it beats RGB and filters. Plan a spectral material classification validation test structures the trial, and if your band question reaches past 1000 nm, start from SWIR vs NIR vs visible imaging.

One note on hybrid dual-band filters that pass visible and NIR onto the same sensor: the red channel then mixes real red with NIR, so any index built on it needs per-unit empirical calibration. That trades accuracy for simplicity, which suits high-volume screening rather than measurement.

Which route fits your case

Your case Route Why it wins
Stopped parts or a static scene Filter swap, two captures One camera; plan 5-10× exposure and a frozen scene
Moving belt, per-part NDVI Beam splitter, two cameras One optical path, simultaneous bands, registration within a few pixels
Flat scene, depth spread ≤ ~5 mm Side-by-side cameras Parallax stays near 4 pixels in the worked example
Canopy with tens of mm of depth spread Beam splitter Side-by-side drifts to about 65 pixels at ΔZ = 30 mm
Moisture, disease, ripeness, material ID Hyperspectral camera Dedicated bands beat any 2-band index

Checks before you order

  • Read the sensor's spectral response curve in the NIR band; where only the visible curve is published, budget the low end of the 10-30% range and add light.
  • Mono sensor for the NIR camera; the Bayer dyes cost response and add crosstalk exactly where you need signal.
  • Global shutter on both cameras of a moving-scene rig, with one trigger line shared by both.
  • Back-focus budget: a 12.7 mm cube consumes about 4.2 mm and each 3 mm of filter glass about 1 mm, from t · (1 - 1/n) at n ≈ 1.5.
  • A lens specified for NIR, or an IR-corrected model; the machine vision lens finder filters by mount, format and resolution.
  • NIR illumination at your chosen wavelength, sized to the exposure budget, with the IEC 62471 rating on file.

To size the rig end to end, start the camera configurator, or browse area scan cameras and hyperspectral cameras. Send the scene speed, working distance and depth spread to the quote form and we will run the parallax and exposure arithmetic with you.

Questions

Can software separate a NIR band from the red channel after conversion?

No. The Bayer dyes mixed NIR into red, green and blue on the sensor itself, and demosaicing interpolated each pixel across its neighbors before your code saw a value. The three split images are near copies of one another, so post-processing cannot re-create a band the optics never separated. The separation has to happen in the optics: a long-pass filter, a filter wheel or a beam splitter, before photons reach the pixels.

Which NIR wavelength should I choose for vegetation work?

850 nm is the default choice. Sensor response is still a usable fraction of the visible peak, long-pass filters and LED floods at that wavelength are commodity items, and leaf reflectance is already near its plateau. At 940 nm the sensor responds less and atmospheric moisture absorbs more light, so the same scene costs more exposure for little index gain. Reserve 940 nm for setups where the illumination must stay invisible.

How do I align the two cameras in a two-camera NDVI rig?

Mount both cameras on one rigid plate with the same lens model, then calibrate: undistort each image and fit a homography that maps the NIR frame onto the visible frame using targets placed across the working volume. That removes distortion and the fixed parallax offset. The depth-dependent shift that survives is e = f · B · ΔZ / Z², and only a shared optical path through a beam splitter eliminates it.

Do I need to refocus after removing the IR-cut filter or fitting an NIR filter?

Yes. Removing glass changes the optical path slightly, and NIR focuses at a different plane than visible light because a lens refracts it less. Refocus with the final filter fitted while the scene is lit only in that band. If one focus must serve both bands in a single pass, look at IR-corrected lenses; otherwise re-set focus between the two captures.

Does the mono version of a camera see more NIR than the color one?

Yes, clearly. A mono sensor has no Bayer dye layer over the pixels, so nothing absorbs part of the NIR before charge is collected and there is no crosstalk to worry about. As a rule the NIR camera in your rig should be mono and the color sensor should take the visible band. The gap is widest at 850 nm and beyond, exactly where a color sensor is weakest.

Is a converted RGB camera good enough for research-grade NDVI?

For relative canopy averages on a static scene, a filter-swapped camera can track changes across a season. It will not deliver research-grade absolute values: the red term is NIR leakage, response varies unit to unit, and the band mismatch drifts with temperature. For absolute indices or per-leaf resolution, use a beam-splitter rig or a multispectral camera and check it against a reference instrument.