I have criticized DxO some posts above (not their numbers, but I questioned their ability to properly interpret them).
Because it is so much easier to criticize than to make a better proposal, I decided to share some bits of a discussion I had a year ago in a German forum, cf.
DigitalFotoNetz.de :: Thema anzeigen - Fomag-Test: Eingangsdynamik und Rauschniveau where I sketched an alternative way of evaluating IQ.
For easier reference, I later fetched my contributions into a PDF (cf. attachment).
Please be aware that this is all in German language and for non speakers of the language, it really isn't worth to try to make sense out of all this
I rather summarize it for you now.
1. State of the art, incl. DxO: noise is considered at the pixel level and measurbated in all possible ways. I.e., noise is only considered at one possible spatial frequency which isn't even typical or standard, but it is the highest possible, the Nyquist frequency.
2. Critics: The assumption that the cross-correlation function for pixel noise is a delta function is naive. On the good part, let's honor DxO for measuring the cross-correlation function for pixel noise. On the bad part, they would just
assume that it is noise reduction applied to raw if they find a deviation. They never ever actually measure a signal ...
My other critics is that it is ridiculous (from a physicist's point of view) that a serious measure for IQ would depend on the
existence of a digital sensor at all. What about IQ for film?
3. My proposition: Have a standard test chart with a
fractal pattern where the signal covers all spatial frequencies. Then have the image mathematically transformed into spatial frequency space and use it to normalize a "constant or full signal".
Now, take photographs of the test pattern, transform, normalize and compare. Take many such images to determine noise. Don't just use the variation from one pixel to another. Of course, a laser projector (or lens clearly outresolving the spatial frequency integration range) must be used.
Then, you get a chart like the following (derived from mathematical considerations):
Blue: Signal curve plotted against spatial frequency (prototypical)
Red: Noise curve plotted against spatial frequency (prototypical)
Green: Quantization curve plotted against spatial frequency (prototypical)
The integrated surface between noise and signal can now be used to define a benchmark for a sensor's (or film's) image quality (at a given ISO step).
Please be aware that I totally ignored color here. But on the positive side, my measure would spot banding, clipping the blacks, or other serious and never measured artifacts.
[Source: all own science
]
When I do criticize DxO, then I mean it at a level they have defined themselves by the amout of effort they claim to put into. And then, I can only say they fail. They fail in a way that you simply cannot seriously say that the sensor with the higher DxOmark is the better sensor. And then, so what?
Last edited by falconeye; 06-15-2011 at 05:29 AM.