Prepared 2014-11-05 by Bill Claff
DxOMark measures noise
for the purpose of determine dynamic range using a well defined testing
protocol.
As part of their presentation they provide Full SNR charts. Theses chart show
Signal to Noise Ratio (SNR) versus gray level (signal).
This data can be easily transformed into a Photon
Transfer Curve (PTC).
A quadratic solve for the variance versus signal should provide measured of
read noise, gain, and fixed pattern noise.
However, it's important to understand that there are many
pitfalls including the fact that the data was not collected with the
construction of a PTC in mind.
In particular, certain forms of noise are not removed from the data; lowering
the SNR values.
Also, the curve provided appears to be a quadratic fit to the original; but the
original data is not provided.
The effects are as follows:
Read noise is largely unaffected because it is the intersect of the fit with the y-axis and the low signal values are not affected.
Gain (e-/DN) will be overestimated. In practice this effect is
most pronounced on the smaller, noisier sensors.
The consequences include inflated values for Full Well Capacity (FWC), Quantum
Efficiency (QE), and read noise (in electrons).
Fixed pattern noise (FPN), also often referred to incorrectly as PhotoResponse NonUniformity (PRNU), is totally unreliable.
Nevertheless, the analysis is a worthwhile undertaking.
Results within certain groups of cameras modes, for example,
DSLRs, should be comparable as they tend to have about the same amount of
inflation.
Linearity with respect to ISO also provides clues that can be interesting about
how a particular camera operates.
In an earlier analysis we already confirmed that light measured at the sensor (Measured
ISO) is linear with respect to the ISO setting.
(There can be important exceptions at the extreme ISO values and detecting
these is useful.)
We expect that gain is directly related to "measured
ISO" and that it will follow the same pattern.
Analysis of the DxOMark data confirms this.
The fits can be viewed on my Gain Chart where
you can see results for multiple cameras.
Because gain has a linear (actually reciprocal) relationship
to ISO it can be characterized as a single value; unity ISO.
This is the ISO value at which there is one electron per DN.
I don't particularly like the term, but this is what it's commonly called.
Full Well Capacity (FWC) at Analog to Digital Converter (ADC) clipping is easily computed from gain.
FWC and "measured ISO" can then be used to compute
Quantum Efficiency (QE).
Note that we're really not computing QE but something more like the product of
QE and fill factor.
You can review these values for multiple cameras on my Sensor Characteristics chart.