Technical question: How does light colour, specifically, affect noise performance?

[Perry Ge]

Here's something I'm curious about but couldn't really find an answer to. Clearly some cameras perform better than others when it comes to high-ISO noise, but the biggest factor in low-noise is, of course, proper exposure.

However, different colours/qualities of light clearly have an effect on noise, as some channels are noisier than others (I say this from what I've read/seen and personal experience reflected in my own photos). I'm wondering about the technical specifics of this kind of thing (and I guess other relevant factors like white balance).

What types, qualities and colours of light are most and least conducive to low-noise?

I figure if anyone can educate me, it'll be someone on this board :D.

[Doug Pardee]

There's no hard and fast answer to this because sensors differ. But in general, it tends to go like this:

1) the green channel is the most sensitive in Raw and also gets the least multiplication into RGB, so it's the least sensitive to noise.
2) the blue channel is the least sensitive in Raw side but needs only modest multiplication into RGB, so it's middling sensitive to noise.
3) the red channel is between the green and blue for sensitivity in Raw but needs considerable multiplication into RGB, so it tends to be the most sensitive to noise.

Lighting that is low on green (or filters that reduce the green) will allow you to increase the exposure without blowing out the green channel, thus reducing the noise in both the red and blue channels. Reddish lighting (tungsten, for instance) or similar filters (orange or red) will allow you to increase the exposure even more without blowing the green and blue channels, thus reducing the noise in the red channel.

Of course you'll either need to do some heavy White Balance adjustments or convert to B&W unless you like off-color photos.

Note: when I refer to "not blowing a channel", I mean avoiding clipping both at the Raw sensor data level and in the RGB (JPEG or whatever) image. The two domains are quite different and it's usually important not to blow any channel in either domain (unless you're going to ignore that channel in a B&W conversion).

[Perry Ge]

Ok awesome! This is exactly the kind of thing I was looking for (I have a 5D btw).

Your description is pretty much how I understand it, but there are two things I don't fully understand.

First, I don't get what 'multiplication into RGB' means.

I also don't fully get the tungsten light part. Tungsten/reddish light is an example of light that is low in the green channel, right? So it sort of restates the point about low-green lighting?

So I take it that it would be worst to underexpose the red channel, then the blue channel, and finally the green channel? And this means that if you are in red-lighting or low-green lighting, underexposure will lead to more noise than underexposure in greener lighting? Or have I got it the wrong way round?

[gjl711]

I don't really think you can equate pixel noise to color. All of the pixels are equal and all generate the same amount electrical noise. The filter over the sensor site may allow less photons through thus increasing the signal to noise ratio, but once through the processing of each photo site will have an identical amount of noise.

However, that said, I would venture a guess that the green photo sites generate more noise as there are twice as many of them as red and blue thus twice the opportunity to miss read and generate a noise pixel.

[Perry Ge]

Woah ok what John said doesn't seem to agree with what Doug said.

My interest and confusion are increasing by the minute!

[Doug Pardee]

All of the pixels are equal and all generate the same amount electrical noise. The filter over the sensor site may allow less photons through thus increasing the signal to noise ratio, but once through the processing of each photo site will have an identical amount of noise.

In a "normal" daylight scene when the green sensor elements are almost maxing out at 4095 (for a 12-bit sensor), the red and blue elements will be ending up with maximum values way below that. Consequently, the signal to noise ratio of green is considerably better than of red or blue.

Now, when we convert to RGB, we need to bring the red and blue values up to be somewhere near the ratios in the original scene as it would appear if lit by the target lighting—usually daylight. Because blue only makes up maybe 7-11% of the final brightness, the blue channel data doesn't have to be multiplied up very much (typically 1/3 to 1/2 stop). But red makes up maybe 20-30% of the final brightness, so the red channel needs to be multiplied up a lot more (typically about a full stop). These multiplications amplify the effect of the noise in those two channels, notably the red channel.

Using reddish light will allow increased exposure without blowing out the green channel, thus increasing the signal in the red channel and reducing the amount of red-channel multiplication needed during WB. That will in turn reduce the amount of noise in the red channel.

Whether it's really enough difference to bother about or not is a different question—one best determined by some experimentation.

[gjl711]

In a "normal" daylight scene when the green sensor elements are almost maxing out at 4095 (for a 12-bit sensor), the red and blue elements will be ending up with maximum values way below that. Consequently, the signal to noise ratio of green is considerably better than of red or blue..... Do you have a reference for this type of noise. I went through the white paper for the 5D (http://www.robgalbraith.com/public_files/Canon_Full-Frame_CMOS_White_Paper.pdf) and the only noise they talk about that can impact image quality is fixed patter noise and random noise. No where do they mention that certain colors are more prone to noise than others.

[Perry Ge]

Well if green is the most sensitive channel because of the higher number of pixels in the bayer arrangement, it does make sense that a properly exposed green channel in daylight will also mean red and blue channels that are less exposed and have to be brought up. Right?