How does the D3 achieve such high ISO?

Started Feb 17, 2008 | Discussions thread
ejmartin Veteran Member • Posts: 6,274
Re: 1D3 full well?

ohyva wrote:

ejmartin wrote:

ohyva wrote:

Second I'm still sceptic about these well capacity "measurements". No
direct measurement method AFIAK (unless you probe into the chip which
I think has not been done (and would probably to some extend
interfere with the masurement). If done indirectly based on noise
figs, then I'd really love to know how all the different noise
sources are separated. And how to ensure there is no in-sensor
digital domain NR applied - which is reported by Sony to be one of
the methods how they have reduced the noise levels in their new CMOS
sensor technology.

There are three common noise sources -- the electronic noise in the
readout and amplification circuits, thermal noise in those same
circuits, and photon shot noise. They are all independent and can
all be measured by varying the exposure.

Are they? I did study for 3 years semiconductor technologies as my
minor in university, and I'd not dare to make this sort of statement

  • but that was now close 3 decades ago when those technologies were

quite young. Perhaps you can point me some more recent technical
studies clarifying that.

Quantitative studies may be found at

And you should not forget the dark current noise which is the
dominant noise source in long exposures.

Dark current noise is thermal noise. In the above referenced sources, average dark current is a small fraction of an electron per second, compared to read noises which are several electrons. So for exposures less than a second or so, thermal noise is totally negligible, as I mentioned.

The thermal noise is proportional to the time of exposure, and can be
measured in a series of longer and longer exposures with the lens cap
on; it is typically irrelevant until exposure times exceed a second
or more.

Have you counted the increase of thermal noise flow with exposure
time due to sensor heating? What anout all sort of paracitic leagages
which may be severy (and non-linearily) dependent on the collected
charge, the trasmitted signal level and of course sensor temperature?

All totally negligible under normal operating environment and exposure settings. See the above references.

Electronic noise is there independent of exposure, it just comes from
reading out the sensor, and so can be measured at very short
exposures with the lens cap on (negligible thermal noise, no photon

I'd not be too sure of this either, as some of the paracitic
components (being the deliver mechanisn on the intra-component
transition noise) in the chip may vary on the signal level - and thus
cause unknown quantities.

I have no idea what you are talking about here. Please be specific, rather than alluding to vague "unkown quantities" look at the data and state an effect or contribution to noise which is unexplained by the standard noise model.

Photon noise varies proportional to the square root of the number of
photons collected, and can be measured by a set of exposures at
higher illumination levels where the other noise sources are
negligible (of course, one can subtract them out for a slight
improvement in the accuracy).

So they can all be separated from one another by varying the
exposure, since they all vary differently with exposure.

Probably true to some accuracy level. But I cannot say if that within
unit of percentages or tens of percentages.

The standard noise model is quite accurate. I've done enough sensor measurements and seen enough of others' analyses to see it verified to the level of a few percent accuracy or better.

As for whether NR is applied, all Canons and probably other CMOS
sensors use a method to cancel noise in the readout of a pixel. This
is not NR as most people think of it, ie smearing out neighboring
pixels to smooth out the noise grain; rather it is done for each
pixel independently and involves no comparison with neighboring

Yep, the coherent doubly sampling can reduce, typically not eliminate
the transition based noise. Or when poorly designed it actually
double the noise level (2nd outcome from the same equations), but I
guess Canon has not made a bad design.

Noise reduction that involves averaging over nearby pixels
is easily detected by looking at the Fourier transform of the image;
if this sort of NR were being performed one would see the noise at
high spatial frequencies take a nose dive relative to lower
frequencies. I have examined raw images of the D3 in this way and
see no evidence of such NR -- the noise spectrum is flat all the way
out to the limits of the sensor resolution (Nyquist frequency).

As NR is typically decision controlled median type of local filtering
I very much doubt you can see that clearly in any fourier based
analysis. But I have to admit it's quite long time since my last math

That is easily checked, I'll do an example and post it if I have the time. But median filtering is a linear averaging over neighboring pixels, and therefore will erase the noise spectrum at high spatial frequency, no question.

-- 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