DxOMark Read Noise Analysis

Prepared 2014-11-09 by Bill Claff


This is my third article on the processing of DxOMark data. It build heavily on the previous: DxOMark Gain Analysis.

Read Noise Model and Fit

Read noise is affected by gain so gain is central to modeling how read noise behaves.

The measured gain is a product of analog gain and digital gain.

gtotal = ganalog * gdigital

Note that digital gain can either be a push ( > 1 ) or a pull ( < 1 ).

Because DxOMark does not gather data at intermediate ISO settings, like ISO 320, I assume digital gain is only applied at high ISO (push) or occasionally at low ISO (pull).

Analog gain can also be accomplished through more than one stage of amplification; two stages are most often used for intermediate ISO.
For the analysis I assume only one stage of analog amplification.

With these assumptions the read noise model is:

ntotal = ( ( (nread * ganalog)2 + (nADC) 2 ) * (gdigital) 2 ) 1/2

Where nread is photosite (pixel) read noise in electrons and nADC is Analog to Digital Converter (ADC) noise in ADUs (DNs).

I introduced the concept of nADC in Sensor Analysis Primer - Read Noise - Optical Black (last paragraph) but can't say I've ever seen it measured and reported.
As we will see, it's a valuable piece of information.

nADC can never reach zero because of quantization error; but it can be quite low.
It's very hard to distinguish from the data between ganalog and gdigital particularly when nADC is low.
So in the initial analysis, gdigital is always assumed to be 1.
This is a mixed blessing since when nADC is not low, gdigital does have an effect on the fit.

Because of these restrictions not all cameras produce a good fit.
Generally poor fits are visually obvious.
Although I'm reluctant to intervene I think I'll need to reintroduce gdigital to the fit and supply the ganalog range in some future analysis.

The fits can be viewed on the Read Noise in ADUs Chart and the Read Noise in Electrons Chart where you can see results for multiple cameras.
I prefer the ADU-based chart to the electron-based one because I think it better reflects what's happening.
But many people prefer the electron-based style, so I have included it as well.

Additional Data

Read noise and ADC noise have been added to the Sensor Characteristics chart.
Remember, to check the validity of a values it's always a good idea to view the appropriate fit on the detailed charts.