Great job from Sony!!! See D3200 review

Started Jul 26, 2012 | Discussions thread
Aku Ankka
Aku Ankka Contributing Member • Posts: 591
Going more and more off topic ;)

ProfHankD wrote:

ProfHankD wrote:

Aku Ankka wrote:

DR increases when noise is reduced. The less there is noise, the more there is DR.

That relationship between DR and S/N is something we teach our EE undergrads....

Are you one of them?

Not for over 3 decades. I'm a professor of electrical and computer engineering. You?

I'm retired

Basically, variants of clipping/median-filtering/posterization can reduce visible noise. Do they improve S/N ratio?

Yes they do.

Ok, let's take the sensor readings and posterize them by a simple threshold to black/white (0 or full scale). Nearly any constant-color patch in the scene will produce pixels with no statistical variation in their values -- which the dpreview analysis would call zero noise! On the other hand, the dpreview analysis would fail to distinguish more than 2 tonal steps, thus showing a very small DR.

Has this thresholding really increased S/N and thus DR?

I would not call posterizing in that way as en efficient way of doing noise reduction Unless the image is essentionally just noise, you're using way too much noise reduction in this kind of sample.

DxOMark missed the NR of ISO 6400+ - a simple FFT shows there is smoothing of the data being implemented at that ISO and beyond (a pity Sony does this), and as you see, it has clear implication to the results.

I am not saying that all noise reduction increases DR - you can smooth everyting a a single value - no more noise, not much information either However, in general NR does increase DR.

Highly correlated noise is still noise, even if the dpreview test -- and human observers -- do not see it as such. In fact, it is the correlations in noise, rather than the noise level itself, that produce most of the perception of noise in images... and that's why noise becomes a complex system effect subject to tweaking....

Well, this if quite a bit OT - first, DPR tests are among the worst to be as examples for noise and/or dynamic range properties of raw data (and/or sensors).

I of course agree with you regarding correlations in noise - pattern noise is very visible and objectionable - and difficult to remove without damaging the image.

Even more importantly, if I didn't agree with you on that I'd be claiming the highlight areas of the images are the noisiest ones for the human perception, which they are of course not, but if we measure the noise, they are the noisiest

Outside of the realm of abtraction everying is noise.

-- hide signature --


Post (hide subjects) Posted by
Keyboard shortcuts:
FForum PPrevious NNext WNext unread UUpvote SSubscribe RReply QQuote BBookmark MMy threads
Color scheme? Blue / Yellow