What Creates Noise/Grain At Higher ISO Speeds?

Olin K. McDaniel wrote:

On Sun, 14 Nov 2004 15:01:36 -0500, Alan Browne
wrote:



Film suffers larger physical grain (dye clouds) in the plane of the film (call
it x,y) as the ISO goes up. Faster speed = larger 'grain' or dye clouds. This
results in the grain that you see. It's not really that cut and dry, but the
end result is coarseness in the negative image that we call "grainy".



Just as a start - I feel it's misleading to compare "grain" in film,
since the above explanation (with which I agree!) shows grain in film
as larger "clumps" BUT noise in digital is quite different. Others
have done a rather good job of explaining noise in layman's terms.
Such as:

Since the OP is asking for a comparison, we can't attribute it to ghosts, phase
of the moon or the price of sardines.

I think the explantion I made (that you clipped) was reasonably clear,
specifically:
"[for digital] the x,y is constant regardless of the sensitivity.." and
"[for digital] The 'grains' remain the same size, but there is a
greater dynamic difference between them due to noise."

Cheers,
Alan

--
-- r.p.e.35mm user resource: http://www.aliasimages.com/rpe35mmur.htm
-- r.p.d.slr-systems: http://www.aliasimages.com/rpdslrsysur.htm
-- [SI gallery]: http://www.pbase.com/shootin
-- [SI rulz]: http://www.aliasimages.com/si/rulz.html
-- e-meil: there's no such thing as a FreeLunch.

Olin K. McDaniel wrote:
In addition there have been lots of other good analogies regarding the
signal to noise ratio effect. My choice takes me back to the early
Radar of WW-II. On the "A Scope" the display was a horizontal sweep
based on distance, and the noise was a random "fuzzy" vertical display
above the noise floor. When a signal came back from the pulsed
output, it showed up as a vertical spike above the noise floor. If
this was weak, the gain was increased, giving a slightly more visible
signal. BUT it also increased the noise level, so the distinction
between the two was just as marginal as before. But it did make the
signal a little more definable or visible.

Physics hasn't changed since WW II and I assure you that the radar problem today
is the same as then. Now we just have much better, more compact, 'faster' and
cooler components which pay off in performance. Which coupled to computers to
perform CFAR, target tracking and 'video processing' (for noise reduction) and
many others ... result in systems that are much easier on the operator.

In fact, for the tracking of airborne targets the problem is much worse than
passive optical imaging as the return signal power is 1/r^4 (r being range) v. a
noise that is constant. Modern radar signal processing has moved the digital
stage ever closer to the receiver portion of the system and this has helped
reduce noise tremendously and in the the case of Doppler navigation radars, has
increased accuracy through the near elimination of tracker bias. (alas, due to
GPS and RLG-INS, Doppler nav is all but obsolete now).


My analogy thus goes like this - increasing the ISO is equivalent to
increasing the gain. It raises the signal, but at the same time it
raises the noise. Then, the camera's "autoexposure" takes over!
Result - some improved image, but much more noise than desired. Thus
we have largely ignored the "autoexposure" effect.

It's not analogy. Increasing the ISO -is- increasing the gain prior to A/D
conversion. There aren't enough bits in the A/D to do all of the gain digitally
without creating massive quantization noise.

Autoexposure has NOTHING to do with the signal itself ... autoexposure takes
place before the image is captured in order to determine the aperture and speed
setting (and the ISO too in fully auto cameras ... ugh).


Bottom line - if the subject is poorly illuminated, the resultant
image is GOING to contain objectionable noise at higher ISO. There
is "no free lunch"!!! Now, maybe in time, technology can improve the
filtration of noise, to improve S/N ratio, but it will take time.

Several noise reduction techniques are employed such as measuring the noise of
the device when no light is hitting it (shutter closed). This noise image can
be subtracted from the image to reduce the noise (though never perfectly).

Of course higher pixel densities lead to smaller photosites ... for which the
noise is somewhat constant, but as the photosite goes down in size, then the
signal drops and consequently the S/N falls. This is the chief failing of ZLR's
and to a lesser extent a flaw in the 4/3 system.

Cheers,
Alan.

--
-- r.p.e.35mm user resource: http://www.aliasimages.com/rpe35mmur.htm
-- r.p.d.slr-systems: http://www.aliasimages.com/rpdslrsysur.htm
-- [SI gallery]: http://www.pbase.com/shootin
-- [SI rulz]: http://www.aliasimages.com/si/rulz.html
-- e-meil: there's no such thing as a FreeLunch.

Ken Tough wrote:

Roger N. Clark wrote:


So, if you want lower noise images, buy a camera with a
larger sensor to collect more photons for a given exposure.


Nice post. Are there cameras or post-processing programs
which can use multiple exposures to average out noise? With
some clever handling, they could keep edges from one image
and create better estimates for colour and intensity by
combining the images together.


Yes there are such programs. Astrostack is free,
ImagesPlus is another (not free, not not real
expensive either). This page shows some results and
how much more you can get out of averaged images with high
signal to noise (99 images averaged):

http://clarkvision.com/astro/saturn.03.02.2004

A lot of my astrophotos are multiple image averages,
e.g. full frame DSLR average of 40+ frames, each
registered to the others:

http://clarkvision.com/galleries/gallery.astrophoto-1

There are others that to much better work than me
with this stuff.

Roger

Ken Tough wrote:

Roger N. Clark wrote:


So, if you want lower noise images, buy a camera with a
larger sensor to collect more photons for a given exposure.


Nice post. Are there cameras or post-processing programs
which can use multiple exposures to average out noise? With
some clever handling, they could keep edges from one image
and create better estimates for colour and intensity by
combining the images together.


Yes there are such programs. Astrostack is free,
ImagesPlus is another (not free, not not real
expensive either). This page shows some results and
how much more you can get out of averaged images with high
signal to noise (99 images averaged):

http://clarkvision.com/astro/saturn.03.02.2004

A lot of my astrophotos are multiple image averages,
e.g. full frame DSLR average of 40+ frames, each
registered to the others:

http://clarkvision.com/galleries/gallery.astrophoto-1

There are others that to much better work than me
with this stuff.

Roger