In part one of this blog (Noise, Photography & Art, Part 1), I mentioned two ways in which noise is injected into the picture in a digital camera. I wanted to conduct an experiment to find out just how much noise each of the two sources provided. As a quick reminder, the first source of noise was the amplifier noise, and the second was the thermal noise in the sensor.
What I did was cover the lens with the lens cap so that no light would get to the sensor. Since some small amount of light can get in through the viewfinder, I covered this as well. So we should have a completely black photograph. I set the camera to raw mode, which means it captures the image with no digital processing having been performed on it, and set the camera to manual so that I could make all of the settings without the camera trying to modify or compensate for anything.
The first sequence of shots I took was with the aperture set to f8 and at 1/125 of a second. Then I adjusted the ISO from its minimum of 100 all the way through to a maximum of 6,400. I examined the resultant file sizes and subtracted the smallest file size from the range. I figure that the difference in file size is a reasonably good measure of the noise. The resultant graph is shown below.
As expected, the amplifier noise starts to increase almost exponentially when we get to the higher ISOs. It is also interesting how we see the saw-tooth pattern with the lower ISOs. I asked around and found that this is actually typical in Canon cameras and is because not all ISO settings are modification to the gain. Some of them use digital processing to pretend that the ISO setting is available. Ironically, this digital processing is injecting more noise into the process, and so this exercise has been very useful to me. I'll only use ISOs of 160 and multiples thereof, unless I want more noise.
So, with that information in hand, I now know that the minimum amplifier noise is at an ISO of 160. For my second experiment I wanted to find out the noise injected by the thermal warming of the sensor. I set the camera again to f8 and an ISO of 160 and this time varied the shutter speed from 1/800 second to 30 seconds. I found almost no noise with the typical shutter speeds and so removed them from the resulting graph shown below.
The noise starts to increase around 0.6 second and then plateaus for a while before starting up on an exponential slope. However, the amount of noise is actually tiny compared to the amplifier noise. In both cases the vertical axis is the file size in kbytes that I attribute to noise, and even the maximum sensor noise is less than the noise injected by the digital processing of the ISO at the best of settings.
Just for complete information, the camera tested is a Canon 60D with an 18M pixel sensor and the base file size was 18069kB.
Perhaps some of you could conduct similar experiments on other cameras. It would be interesting to see how a Nikon compares. Also, if anyone has any information of the ISO processing being done in these cameras I would be very interested.