When to choose a Bayer color camera: a decision checklist

Most color machine-vision jobs do not need a prism camera. A Bayer color camera measures one channel at every pixel and reconstructs the other two by interpolation, which costs per-pixel color truth, some signal and some edge sharpness — and nothing that a sort, count or presence check notices. This checklist is for the engineer specifying a color camera for an inspection line. Compute the pixel budget with the Bayer allowance, classify the color task as relative or absolute, then walk the grouped checks: the decision falls out of two numbers and a handful of lines. Bayer wins for relative color with a workable pixel budget; a three-sensor prism waits for absolute color matching or true per-pixel RGB.

What a Bayer camera actually captures

A Bayer sensor places a mosaic of red, green and blue filters over the pixels: two green, one red and one blue in every 2×2 block. Each pixel measures exactly one channel; the two missing values at that site are filled in by demosaicing, an interpolation step that estimates them from neighboring pixels. The frame looks full-color, but only one color value in three is a measurement — the other two are estimates.

Three consequences drive every check below. Interpolated color is consistent across a frame, so region-to-region comparison works, but an RGB triple from a Bayer camera reflects the reconstruction algorithm as much as the scene, so matching an external reference does not. Luma detail rides mostly on the green channel, which samples at half density along each axis, so treat usable geometric detail as roughly half the pixel count. And each pixel receives only its own band, so the same sensor in mono gathers roughly three times more light per pixel. The color filter array explained: monochrome versus Bayer sensors guide covers the mosaic itself; this page is the buying decision.

The two checks that decide it

Check one, the pixel budget:

FOV is the field width in mm and r the smallest feature you must resolve or the position tolerance, also in mm. n is the mono-equivalent pixel count across the field. The doubling is a working rule of thumb for interpolated luma: the green channel samples half density per axis, and the reconstruction spreads red and blue over neighbors. Any color sensor with at least n_Bayer pixels across the field passes. If none does, buy more Bayer pixels — that costs less than any prism route — and see color camera spatial resolution: beyond the pixel count for how a filter mosaic interacts with measured detail.

Check two, the color task. Relative color — is this cap the same class as those, is the gasket present, is this the front face — passes on Bayer. Absolute color — matching a master chip, a lab-measured RGB triple or another plant's readings — does not. The identify part front and back by color differences guide is a typical relative task; compare color between two images with meaningful metrics shows what a genuine comparison needs.

Decision rule: Bayer passes when the color task is relative and the sensor clears n_Bayer. A prism camera enters only when one check fails for a reason pixels cannot fix.

Worked example: colored caps on a moving line

The job: sort 30 mm caps into red, blue and black, and hand cap centres to a pick arm at ±0.15 mm. Lane wander sets the field of view at 60 mm.

Pixel budget: n = 60 / 0.15 = 400 px across. Bayer allowance: n_Bayer = 2 × 400 = 800 px. Typical 5 MP sensors span about 2,448-2,592 px horizontally, so the margin is 2,448 / 800 ≈ 3.1 — roughly three times the requirement.

Color task: the classes are learned from caps imaged under the same light on the same line, and the classifier compares region means — relative color, so reconstruction errors cancel. Geometry: the centre comes from a centroid over thousands of pixels inside the cap boundary, not from a sub-pixel edge fit, so smoothing at the rim cannot move the mean materially.

Decision: a 5 MP global-shutter color camera does the whole job. Run on-device, the MV-SC3050XC-16M-ATN-H covers it: 5 MP color vision sensor, global shutter, 1/1.45" format, M12 mount, Fast Ethernet output.

The same arithmetic scales. Tighten the tolerance to 0.05 mm on the same 60 mm field and the budget becomes n = 60 / 0.05 = 1,200, so n_Bayer = 2 × 1,200 = 2,400 px across. That still fits a 12 MP color frame. The MV-CH120-11UC is a 12 MP color camera on USB3 with a global shutter, so the margin stays comfortable. Prism never enters this example at any tolerance, because the color task stays relative.

The pre-commit checklist

Run the items in order and stop at the first failure.

Color task

• Color separates or flags parts — sort, presence, front/back — rather than grading a shade. Bayer is sufficient. Why: interpolation shifts absolute values but preserves the separation between regions imaged together.
• References are learned on the same camera under the same light. Fine on Bayer. Why: any systematic color error applies to reference and part alike and cancels.
• The test compares against an absolute standard — a master chip, a lab RGB triple, another site's values. Use a prism camera, or calibrate the Bayer camera against that standard in place. Why: demosaiced RGB tracks the reconstruction algorithm as much as the scene.

Spatial budget

• Compute n = FOV / r and double it. Why: the doubled figure is the sensor pixel count you need across the field, per the rule above.
• Keep about 2× margin above it. Why: the doubling is optimistic at chroma edges, and lens falloff and focus drift take more.
• No sub-pixel edge measurement from the color image; if geometry demands it, add a mono measurement camera and keep Bayer for classification.

Light and motion

• Each Bayer pixel sees one band only, so add light before you lengthen exposure. Why: two of the three bands are absorbed at every pixel — the three-to-one signal ratio above.
• Parts move during exposure: keep blur under a third of the smallest feature, t = 0.33 × r / v — a 2 mm feature at 300 mm/s gives t = 0.33 × 2 / 300 ≈ 2.2 ms — and pick a global shutter; rolling shutter shears fast parts at any color technology. For fast lines, a global-shutter color model such as the MV-CU020-90GC, 2.5 MP over GigE, is the standard pick.

Integration and ordering

• Plan the pipeline around raw Bayer: one value per pixel, mono-like bandwidth at equal resolution. Grab the raw mosaic when you want to control the reconstruction yourself, and step to 10- or 12-bit modes when shades sit close — raw10 and raw12 sensor modes: get full pixel depth instead of raw8 covers the modes.
• Order the Color variant, not Mono. The same camera families ship both under near-identical model numbers; the catalog's colour filter line is the reliable tell, and the wrong variant is an expensive discovery at commissioning.

When Bayer is the wrong answer

• Absolute color matching: brand-color QC against a master sample, colorimetry, cross-site comparison. The reconstruction, not the scene, defines the values; use a prism camera. The prism 3-sensor color cameras vs Bayer: when true RGB matters comparison lays out where true per-pixel RGB earns its keep.
• Sub-pixel geometry measured from the color image: reconstruction smoothing caps edge precision, so move measurement to a mono camera or to prism.
• Subtle shade grading: batch tint differences a few RGB counts wide sit under the interpolation error and the lower signal; prism separates them.
• No hue information at all: buy mono. A colored light tuned to the contrast you need often separates features better than a color camera under white light, with more signal per pixel.
• The overspend case in reverse: when the doubled budget outruns the sensor you had in mind, buy more Bayer pixels before you buy prism. The MV-CS200-10GC carries 20 MP of Bayer color over GigE on a 1" sensor — headroom that clears almost any pixel budget from the checks above.

Bayer vs prism at a glance

What to do next

Short version: relative color plus a cleared pixel budget is a Bayer camera; absolute color or per-pixel RGB is a prism. Filter the area scan cameras to Color and your interface, or step to smart cameras & vision sensors when the inspection runs on-device. The camera configurator walks sensor, mount and interface in order, and the camera finder sorts the catalog the same way. If the two checks leave a live line undecided, contact us with your field of view, tolerance and part colors — the arithmetic takes minutes.

More camera-side guides sit on the cameras topic page.