Prism 3-sensor color cameras vs Bayer: when true RGB matters

Most color inspections never need a prism. A single sensor with a Bayer color filter array measures one channel at each pixel and computes the other two, and when your color patches are a hundred pixels wide that estimate holds up. This comparison answers where it stops holding up: shade grading to tight tolerances, LED panel inspection, parts that mix deep red varnish with pale blue print, color features only a few pixels across. You get the number that separates the two architectures, a worked carton example with focal length, working distance, exposure and f-number arithmetic, and a decision for each task. One commitment up front: if your lighting drifts, no camera architecture saves you — stabilize the light before you buy anything.

The decision, up front

Buy a prism 3-sensor camera when one of three conditions holds. First, the smallest feature whose color you judge spans fewer than four object pixels — the Bayer unit cell is 2×2, and inside that cell color is inference, not measurement. Second, the part's brightest and dimmest channels sit more than about three stops apart (a 10:1 red-to-blue ratio is 3.3 stops) and your lighting cannot even them out. Third, channel registration itself is the metric — display pixel pitch, fiber color, hairline color bars. In those three cases measured beats estimated, full stop. For sorting, presence checks and print verification on millimeter-scale patches, a Bayer color camera wins on resolution, size and simplicity, and the rest of this comparison keeps its errors out of your tolerance.

What a Bayer sensor really measures

Half the photosites carry green filters, a quarter red, a quarter blue. Each pixel reads out one measured channel; the other two are interpolated from its neighbors. Two of every three RGB values in the file are computed: a 12 MP color camera measures 12 million single-channel samples and derives 24 million channel values from them. Luminance detail survives, because green is sampled at every other site in each direction. Chroma detail does not: red and blue hold a quarter of the sites, so effective color resolution is about half the pixel count in green and a quarter in red and blue — the limit behind color camera spatial resolution.

The failures are predictable. Neutral edges grow one- to two-pixel zipper fringes, and repeating patterns near the red/blue sampling limit produce color moiré that no decoder removes, because the values were never captured. Quad-Bayer sensors meet the same wall from the other side: bin for sensitivity, remosaic back to an estimate.

What a prism 3-sensor camera measures

A dichroic beam-splitter prism behind the lens divides the image into red, green and blue bands and focuses each onto its own monochrome sensor — the classic three-sensor camera design. All three sensors expose the same instant, full frame and full resolution, factory-registered to a fraction of a pixel. Every pixel's RGB is measured; nothing is demosaiced. And because each channel is a physical sensor, each gets its own exposure time and gain.

The cost is physical too. One lens must feed three sensors through glass, so the head needs a larger image circle, a long back focal distance and apochromatic-grade color correction. It is bigger and heavier, it outputs three full frames per image, and resolution sits lower: prism heads on the market cluster in the 2–10 MP class, while single-sensor color ships at 20 MP and beyond.

The numbers side by side

Per-channel exposure is worth 3.3 stops

Take a warm-white LED lighting a part with deep red varnish and pale blue print. The red channel collects ten times the blue. One exposure must fit red without clipping, so blue digitizes at a tenth of its scale: log2(10) = 3.3 stops down, about 25 of 255 gray levels on an 8-bit sensor. Raising blue gain afterwards multiplies read noise tenfold — gain cannot manufacture photons. A prism camera integrates blue 3.3 stops longer than red, collects real photoelectrons in both, and reads both near full scale.

LED panel inspection is the textbook case: each display pixel may span one or two camera pixels — inside the Bayer cell — and blue diodes run dimmer than green, wanting their own exposure. Mind the panel's own PWM dimming too, which can beat against your frame timing; the mechanics of RGB LED dimming, color shift and flicker are worth reading before you fix a frame rate.

Worked example: carton shade and print on a belt

Cartons 55 mm wide on a 150 mm/s belt, shade judged on 2 mm patches, carton height varying ±1 mm. Field of view with handling margin: 60 mm.

The camera: the MV-CH120-11UC, a 12 MP, 1.1", USB3 color camera with the IMX304 sensor and a global shutter. A 1.1" sensor runs about 14 mm wide, and 12 MP in 4:3 puts roughly 4,000 pixels across — 14 mm / 4,000 = 3.5 µm pixels.

• Magnification: m = 14 mm / 60 mm = 0.23×
• Object pixel: 60 mm / 4,000 px = 15 µm
• Shade patch: 2,000 µm / 15 µm ≈ 133 px — 33× the 4 px Bayer cell, so this color is measured well enough to trust

The lens: the LS-FFL-0567, a 16 mm, 1.1", 20 MP C-mount machine vision lens.

At a 16 mm focal length the working distance is 16 × 1.23 / 0.23 ≈ 85 mm. A 1.1" image circle on a 1.1" sensor leaves zero margin, so for margin or standoff use the LS-FFL-1676: 60 mm focal length, 65 MP, M58, rated 214–1536 mm working distance. That glass holds the same field at about 320 mm.

Exposure: half a pixel of blur at 150 mm/s is 7.5 µm, so t ≤ 7.5 µm / 150 mm/s = 50 µs. Run a 50 µs strobe; the global shutter keeps it clean. Belt speed, blur budgets and strobing are worked through in sorting conveyor camera selection.

f-number: for ±1 mm of height change allowing one pixel (3.5 µm) of defocus blur, DOF ≈ 2 × f# × c × (1 + m) / m² = 2 × 5.6 × 0.0035 × 1.23 / 0.0529 ≈ 0.9 mm at f/5.6. Stopping to f/8 buys 1.3 mm, but the Airy disc alone is then 2.44 × 0.55 µm × 8 ≈ 11 µm — three pixels of blur before the part arrives. Small pixels want f/4–5.6 and mechanical flatness, a trade worked through in small-pixel camera selection. Commit: fixture the cartons flat within ±0.5 mm and run f/5.6.

And the prism crossover for this exact set-up: 4 px × 15 µm = 60 µm. If the brand microdot on that carton is 50 µm across and its hue is pass/fail, it spans about 3 px — inside the cell. Below 60 µm at this field of view, buy magnification or a prism head.

What a prism head asks of lens and grabber

The lens first. The prism sits between the glass and the sensors, so the head needs a long back focal distance, frequently a mount beyond C-mount, and an image circle over the largest chip. Chromatic correction is the hard number: lateral color of 0.1% across a 14 mm width is 14 µm of red-versus-blue shift — four pixels landing as permanent channel misregistration that no exposure setting repairs. That is why prism heads carry apochromatic-grade glass. For lens choice in general, machine vision lens selection and the format-matching rules in lens sensor format explained apply to both architectures.

Bandwidth second. Three channels triple the payload: 12 MP at 12 bits per channel and 30 fps is 12 million × 36 bits × 30 ≈ 13 Gbps — 10GigE or CoaXPress country — so budget one of our frame grabbers with the head. The Bayer example above carries 12 million × 8 bits × 30 fps ≈ 2.9 Gbps, comfortable on a USB3 cable. Interface and pixel-format ceilings are laid out in RGB camera frame rate limits.

Treat the head as a metrology instrument. The three sensors are registered at the factory to a fraction of a pixel, and hard shocks or large thermal swings can move that: mount it soft, recalibrate after any drop.

Checks before you order

• Measure the smallest color feature in object units, divide by the object pixel size, and apply the 4 px rule before anything else.
• Image your worst part, not your golden sample: fine print, sharp color borders, glossy curves. Hunt zipper fringes and moiré at 1:1 zoom.
• Match the lens image circle to the sensor format with margin; for a prism head, confirm flange distance and the apochromatic spec in writing.
• Match shutter to motion: global for strobed belts, rolling only for static scenes.
• Read the light's spectrum, not its brightness: white LEDs with a weak deep-red tail starve the red channel, and gain cannot fix what the source never delivered. A broad-spectrum source or a red booster channel does.
• Compute bits × pixels × frames per second against your interface before you pick the cable, NIC or grabber.
• Calibrate against a color chart on day one and re-check weekly; log ΔE when shade is the pass/fail call.

On the Bayer side, three honest paths through our color area scan cameras:

Values as printed in each product's data sheet; open a part for its drawing and full specifications.

The MV-CS200-10GC (20 MP, 1", C-mount, GigE, IMX183, rolling shutter) covers large static fields; the MV-CH120-11UC above is the moving-parts pick; the LS-CMR-0053 (12 MP, 1/1.7", C-mount, GigE, IMX226, 9 fps) suits fixed stations with controlled light.

When the usual answer is wrong

The overbuy: a metrology-grade prism head watching lighting that drifts 5% over a shift. That drift swamps the camera's accuracy; a Bayer camera with chart calibration and stabilized, broad-spectrum lighting outperforms an accurate camera fed by an unstable one. The lighting side of that fight is covered in color sorting under changing ambient light.

The underbuy: textile weave, halftone print and laser-etched micro-text sit right at the red/blue sampling limit and demosaic into false colors your algorithm then chases. The data was never captured; no decoder recovers it. Use a prism head, or defocus slightly to low-pass the scene.

The misuse: per-channel exposure stretched across three aging LED strings to hide their differences. Exposure per channel belongs to the part's physics, not the light's decay — replace the strings.

The ceiling: needing 20 MP and up in color puts you in single-sensor country today — for example the MV-CS200-10GC at 20 MP, or the 25 MP, APS-H, F-mount MV-CH250-25TC-F-NF. Accept the demosaic, or split the field across two cameras.

The trap: a rolling-shutter Bayer on a fast belt records each row at a different instant, so color edges smear even when the frame rate looks sufficient. Global shutter or strobe — never paper over it with software.

The short version

• Color patch spans 4 object pixels or more, light stable: Bayer color camera, chart-calibrated. Done.
• Patch under 4 px, or registration is the metric: prism 3-sensor.
• Brightest and dimmest channels more than 3 stops apart, lighting cannot fix it: prism, for per-channel exposure.
• Repeating fine color pattern — weave, halftone, micro-text: prism, or defocus.
• Resolution above 20 MP: Bayer, accept the demosaic.
• Light drifting: fix the light before buying either architecture.

Start with the area scan camera finder for the Bayer side, or open the configurator at the camera step and work lens, light and controller from there.