Here is a link to an interesting overview of digital camera colour detection methods:
http://thedailynathan.com/ee119/Yan,%20Nathan%20-%20An%20Overview%20of%20Col...%20Sensors.pdf
Three things in particular caught my attention:
1. A colour filter element in a Bayer CFA either absorbs or reflects light outside its pass-band. It also partially absorbs light within its pass-band. These two factors mean that using a Bayer CFA in front of a standard sensor, which is intrinsicly a monochromatic device, reduces its overall efficiency to 20% of an unfiltered sensor - that's -2.3 stops light loss (p.7 & footnote 14).
This loss is from the CFA alone. There are other transmission losses from the AA & IR filters and from the quantum efficiency of the sensor.
2. Quoting from p.3:
There is however, one key issue with the functionality of this basic photodetector – it is only able to detect luminance information. An inherent limitation of counting electrons produced by the photoelectric effect for image detection purposes is that chrominance information is lost. For a given electronic image sensor, any photon with energy exceeding (or corresponding wavelength below) the work function of the surface material is able to eject an electron, which can then be counted. The color information of the photon is encoded in its wavelength, or energy, and this is translated to the ejected photoelectron as kinetic energy. Current electronic image sensors simply count the number of accumulated electrons, however, irrespective of their energy, so this chrominance information is lost, and the electronic image sensor is inherently monochromatic – it measures the intensity of light over only a single wavelength range encompassing the visible spectrum, and thus is unable to derive anything about the spectral distribution of the input light within it.
What about if the detector was able to respond in some way to the kinetic energy of the released electrons? Then instead of being a simple monochromatic counter of released electrons (and thus photons), it could determine the colour without a CFA. As to how you would then determine light level (currently done by the number of electrons that arrive in the photowell per sampling period), I have no idea.
3. In the Kodak RGBW filter, one of the G filters in an standard RGBG array is omitted, resulting in improved luminance sensitivity (+50%) and lower chrominance resolution (-25%). Since the eye is not very sensitive to chrominance resolution, the trade-off can still be worthwhile. But an unfiltered sensel saturates more easily than the other 3 filtered sensels in the sub-group, so the RGBW method is best for low-light colour sensors where you're not operating at base ISO sensitivity, but are using significant analogue gain between the sensor and the ADC. (pp.21-26)
Dan.
Last edited by dosdan; 10-08-2011 at 02:46 PM.