4-color Bayer filter arrays: why?

[bg2b]

I have an Olympus E-1 and sometimes convert the RAW images using dcraw.  I've found that I often get "maze" artifacts unless I use the four-color interpolation switch (-f).  To try to understand why, I took pictures of various uniform colored surfaces and looked at the pixel values in the RAW files.  For some reason, it appears that the E-1 has two slightly different types of green pixels.  And from various other suggestions that I've found around the web (such as notes about the recent Lightroom bug with certain classes of cameras on non-Intel Macs) I get the impression that the E-1 isn't alone in this.

My question is why would companies do this?  It's certainly not simpler to make half the green pixels of a different type.  It seems like many demosaicing algorithms can do a better job (resolution-wise at least) if the green response is uniform.  There must be some counterbalancing benefit, but what?  How do you even turn the 4-color data into a 3-color output file?  From looking at dcraw's source, it looks to be just averaging the two greens when you use -f, which strikes me as making the whole thing even more pointless.

Also, when I look at the dxomark data for various cameras that I suspect of having 4-color arrays, they only give information about three color filter responses.  Are they ignoring half of the green pixels?  Am I just confused about the whole thing?

[fredjeang]

Hi,
I had an Olympus E1 too when I restart photography in digital age. This is a very good question that I'm unable to understand, I'm totally useless in technical engineering but I hope some advices people will be able to bring some light on this matter.
The E1 is one of the very few dslr I regret, even in today's 2010 technology. Strangely, this was a camera much more capable than it's reputation and the files from the E1 were just beautifull. The point you are making is a mystery but might have a logical explaination that escapes to my knowledge.
Hope some members will be able to explain this fact.

Regards,

Fred.

[Graeme Nattress]

I could see some logic in deliberately altering the response to be used with a special demosaic, but usually even just a small pixel response difference from how the chip is read out is enough to cause a variation. Any decent demosaic algorithm should be intolerant to such differences though, as they're very common.

Graeme

[ejmartin]

I suspect that the logic is as follows: The two green channels in cameras such as the Olympus models have slightly different spectral responses in the color filters, but overall they are more similar than different.  Averaging the two gives the local G response, taking the difference gives additional spectral information to do better color separation.   Dcraw apparently disregards this extra information, but in principle one should be able to construct a color profile that does.  I don't know whether any converters other than the one from Olympus does, however (I presume Olympus makes use of this information somewhere along the line, otherwise it is indeed rather pointless).

I would agree, though, that resolution must take somewhat of a hit, since most good demosaic algorithms use the higher sampling of the G channel to bring the demosaic'd data's resolution close to the Nyquist of the sensor array; by making the two greens slightly different in spectral response, one cannot tell whether differing green values are due to Nyquist textured luminance information, or aliased color difference information due to the different responses.  Averaging the two (as in dcraw) would erase the highest frequency detail.

[bg2b]

A little more experimenting...

This comes from photographing solid red, green, and blue patterns on my monitor and then looking at differences between the RAW pixel values for green1 and green2 within each 2x2 block of pixels.  Here's a histogram of green1-green2:

[attachment=20870:e1.png]

The red histogram corresponds to the difference in response of the greens when photographing the red pattern, etc.

You can see that none of the histograms is centered at zero, and the fact that the difference in red response goes the opposite way from the difference in blue response says that this isn't just some sort of A/D converter mismatch.  Not that I understand it, but I think it is real.

I still don't know what to make of dxomark for any cameras like this.  It seems like their color sensitivity, color response, and full CS data assume three-color sensors.