Olympus-OM
[Top] [All Lists]

Re: [OM] New E-1 (noise reduction)

Subject: Re: [OM] New E-1 (noise reduction)
From: Joe Gwinn <joegwinn@xxxxxxxxxxx>
Date: Mon, 22 Sep 2003 08:56:19 -0400
At 7:40 AM +0000 9/22/03, olympus-digest wrote:
>Date: Sun, 21 Sep 2003 20:50:54 -0400
>From: W Shumaker <om4t@xxxxxxxx>
>Subject: Re: [OM] New E-1 (noise reduction)
>
>Maybe I'm a cynic. Having been an analog CMOS IC designer for years,
>familiar with all the A/D linearization and calibration schemes. The ultimate
>limit is the core electronics. I look at places like dpreview, which attempt to
>be as objective as possible, but they measure and compare camera noise
>based on a constant area of tone; and I know there are a lot of post processing
>techniques that can be done, so the numbers don't mean much. Consider,
>for instance, the disclaimers that said the pre-production olympus E-1s did
>not have the full noise reduction algorithms working and hence were not 
>valid in comparison to other cameras. Just implies that one can tweak the
>camera software to improve whatever you look at, so the numbers just end
>up being engineering tradeoffs between various algorithms.

This is only a problem if the camera manufacturers have installed special 
software to detect and oversmooth flat areas, on the theory that only in flat 
areas is noise all that visible.  This is certainly possible, but is also 
detectable by the right test.


>Like I said, if one could look at the purely raw electronic data and see
>what the underlying electronics can actually do before any processing,
>maybe a realistic comparison could be made between cameras. I find
>it hard to believe that a camera with a pixel sensor half the size of another
>camera could have the same or better noise performance, unless there was
>some other performance tradeoffs being made. So one camera has higher
>noise but a sharper image. You don't really know if you are just seeing the
>difference in post-processing algorithms.

I don't think it's impossible to test such cameras, even if the internal raw 
A/D counts are not available.  One test occurs to me right off:  With camera on 
a tripod, take multiple picutures of the same scene.  Align the images (if 
needed) and subtract them, pixel by pixel.  What's left is twice the noise 
power in the individual images.  Do this with a collection of scene types, 
including ones with large flat areas.  This is a classic black-box test, is 
directly relevant to practical use, and cannot be evaded, no matter what the 
camera manufacturer does.  

I know it sounds odd that subtraction of images yielded a doubling of noise 
power, but that is correct, a consequence of the conservation of energy.  The 
key is that the random noise is by definition different in the two images, so 
it does not cancel out.  Only the non-random stuff (image plus fixed-pattern 
noise) cancels out.

The full strength version of this test is to process the images as follows:  
For each aligned pairs of images, compute the sum and the difference images.  
Divide each pixel in the difference image by the square root of two, and in the 
sum image by two.  Then, divide each (reduced) pixel in the difference image by 
the corresponding (reduced) pixel in the sum image, yielding the per-pixel 
standard deviation of the image's random noise, expressed as a fraction of the 
average value of each pixel.

Fixed-pattern noise is measured by taking pictures of total blackness.


>OK, I've had my rant. Crank those babies up to iso 1600 and see what they
>can really do.

Yes.  On, to the barricades!


>At 12:45 PM 9/21/2003, you wrote:
> ><snip>
> >A good measurement is to photograph a uniformly-lit surface (which need not 
> >be gray) and compute the average and standard deviation of the values of the 
> >pixels in a patch containing at least 50 pixels.  Do this computation three 
> >times, once per color.  The larger the ratio of the standard deviation to 
> >the average the greater the noisiness of the image.  This is a standard way 
> >to measure such things.
>
>Like I said, this is exactly the kind of test that can be manipulated most by 
>post
>processing in the camera. You have no idea what happened to any potential
>detail that would have existed in a real photograph over that same area. If 
>there
>is no detail, any camera can average out the noise and you would not know the
>difference. We need a noise test that contains detail also, like a real 
>photograph.

As Moose mentioned, we need not and should not care what goes on within the 
camera.  We can make perfectly adequate black-body tests without opening the 
camera up.  Not that knowing how such cameras work won't help us to devise 
devious tests that expose the shortcuts.


Joe Gwinn


< This message was delivered via the Olympus Mailing List >
< For questions, mailto:owner-olympus@xxxxxxxxxxxxxxx >
< Web Page: http://Zuiko.sls.bc.ca/swright/olympuslist.html >




<Prev in Thread] Current Thread [Next in Thread>
Sponsored by Tako
Impressum | Datenschutz