The default answer in machine vision is monochrome, and for most applications it is correct. A monochrome sensor has no colour filter array, so every pixel receives light from the full spectrum - roughly three times the signal of a comparable colour pixel, and a correspondingly better noise floor. Before defaulting to it, though, it is worth checking whether the defect you are chasing is defined by colour.
What a colour camera actually measures
A colour sensor places a Bayer pattern of red, green and blue filters over the pixel array, then reconstructs full colour per pixel by interpolation. That has two consequences. First, effective spatial resolution per colour channel is lower than the pixel count suggests, because each pixel measures only one colour and the other two are inferred from neighbours. Second, the colour filters absorb light, so a colour camera at the same exposure sees a dimmer, noisier image.
The practical impact depends on the defect. A 0.5 mm scratch whose contrast is a brightness change is better detected in monochrome. A 0.5 mm contamination whose only signature is that it is brown on a grey part is invisible in monochrome and trivial in colour.
Cases where colour clearly wins
Material or coating mix-ups. If two part variants differ only by an anodising shade, a wire insulation colour or a label tint, colour is the only way to sort them.
Contamination and oxidation. Rust, copper oxide, oil staining and food residue all read as colour shifts rather than brightness shifts.
Product grading. Agricultural and food inspection grades by hue - ripeness, browning, char - where the discriminant is inherently chromatic.
Assembly verification with colour-coded components. Wire harnesses, cable looms and connectors where position is confirmed by colour sequence.
Human-facing documentation. If images are archived for operators or customers, colour is worth the bandwidth for comprehension alone.
Getting more out of colour when signal is scarce
Do not simply accept the threefold light penalty. Three standard techniques recover most of it:
Use colour LED illumination matched to the feature instead of white light. If the defect is a red mark on a steel part, red light makes it near-invisible and a blue channel view makes it dark - the same sensor, far better contrast, and no filter loss.
Compute on a single channel, or on a channel difference. A monochrome camera with a coloured light performs identically to the blue channel of a colour camera at a third of the data rate. If your algorithm only ever uses one channel, buy a mono camera and a coloured light.
Use 2x2 binning when the colour decision is regional rather than pixel-level, trading resolution for signal.
Choosing the sensor
Compare like with like inside one family. The MV-CS050-10GC and the MV-CS050-10GM share the same 5 MP sensor and the same GigE interface; the GC model trades light and resolution for chromatic information and nothing else changes in the integration. For higher resolution colour work, the MV-CU060-10GM class covers 6 MP in a compact body, and where the inspection algorithm runs on the camera rather than on a host, an embedded colour smart camera such as the MV-SC3016C-06M-WBN removes the host entirely.
The decision rule that holds up in practice: choose monochrome unless the defect you are detecting cannot be described without the word colour. When it can, choose colour and recover the light with coloured illumination rather than with a bigger shutter.