Interesting paper on DSLR noise

[frugal]

I was Googling for information about ISO levels on Canon DSLRs, looking to see some evidence as to which one to use. One of the papers I stumbled across was this one from a physics professor at Chicago University: http://theory.uchicago.edu/~ejm/pix/20d/tests/noise/index.html

The items of interest I pulled out of the article:

1. The voltage pulled of of the CMOS can actually be negative, so Canon add 1024 to the value pulled off of a 14 bit ADC to counter this. Apparently Nikon do not do this, so they actually read negative voltages as 0.
2. The act of reading the signal adds electrons to certain wells, which is where the Bias signal comes from. It is determined by the electrical layout of the CMOS.
3. The conversion from electrons in the CMOS to a digital value is done by the ADC (analog to digital converter). Okay, I knew that
4. The ISO gain factor is applied during by the ADC, which means it is an analog process, not a digital one. As it is an analog process it introduces noise
5. Canon cameras have the "standard" ISOs of 100, 200, 400 etc, but also intermediate ISOs of 125, 160, 240, 320 etc. These are done by a second stage of the ADC that introduces a second bit of noise.
6. ISO100 gives a lower digital value than the actual number of photons received, but the rest give a higher value than the number of photons, so don't use ISO100

Ultimately the advice is to use the highest standard ISO level you can without white clipping as this will give the lowest ADC ISO gain noise, whereas the rest of the noise scales with ISO.

[nmoushon]

So basically its saying use high ISO on dim objects and low ISO on bright object (except dont use ISO100)?

Obviously scaling so that you're not clipping data and taking local LP into that as well.

[Steve Ward]

Interesting read , quite dated now though.

[StuartJPP]

Although I have no expertise in this field, I believe that the quantization error mentioned is what allows us to get away with using higher ISOs (to a point) to tease out fainter detail, with the negative aspect of blowing out the highlights. I have always maintained that the amplification of the signal is in the analog domain prior to digitisation and is dependant on the ISO setting. The difference is subtle since amplifying the signal only really affects signals that are just within or above the noise floor since larger signals would already be detected without amplification, but it allows us to retain this "noise" rather than discard it. As a consequence the image gets grainier/noisier but the noise contains some useful data that when stacked and averaged becomes apparent.

I don't know how the above can be proved/measured since by the time the RAW file is produced the data mentioned has already been discarded.

In a perfect DSLR world each and every photon detected would be converted into a measurable value but unfortunately it isn't.

Disclaimer...Everything above could be complete and utter rubbish as I am no expert.

[frugal]

So basically its saying use high ISO on dim objects and low ISO on bright object (except dont use ISO100)?

Or any ISO that is not one of the "standard ones", as they use two separate sets of scaling.

[dph1nm]

Rather confusing plotting everything in ADU not photons (or electrons). For instance the x-axis on Fig 12 represents different numbers of photons for different ISO, so you can't compare the curves with one another, and the read noise in photons would actually decrease with increasing ISO (fig 14).

Note that some of the results are specific to the particular camera model, so my 1000D has a bias of 256 not 1024 and a unity gain somewhere between ISO 400 and 800 I think , not 100).

NigelM