Noise Reduction Software - please read

Started Apr 3, 2003 | Discussions thread
Uwe Steinmueller
Uwe Steinmueller Veteran Member • Posts: 3,414
Re: Noise removal is best before CCD demosaic process


I think along these lines and some manufacturers probably do something like this in their firmware.

You posting confirms that raw conversion can be improved in the future.


Stuart Nixon wrote:

mike rivers wrote:

With all the info about the noise on the 14n, here is a website to
carefully look at ( the info here can be applied to all digital

Thanks to ResIpsaLoquitur in previous post for pointing out the
Neat Image software

Thanks for an interesting set of reviews.

I've been doing a lot of research on noise removal recently, as
part of some photo editing software I'm working on.

There are a couple of interesting points that are worth flagging
with respect to noise removal (note I'm not commenting on the
process flow rather than on particular software products - many do
excellent work).

1. First, and the primary point:
Noise removal is best done before any other processing.
Specifically, it should be performed before the demosaicing process
(which interpolates Bayer style CCDs back into full RGB values).

The reason for this is that noise is very often color channel
specific. It is a lot easier, for example, to remove a blue noise
value before it has been used as input to generate other
blue/green/red values. In other words, noise in a CCD value will
"spread" to other colors and other pixels during the demosaic
interpolation process. So if you can remove the noise before
interpolation, the result is much better. Noise removal can be more
aggressive, while getting a sharper image out. The disadvantage is
that you have to regenerate the demosaiced image whenever noise
reduction filter values are changed.

There is typically a very poor noise correlation between channels
(e.g. noise at a pixel in the red channel won't be heavily
correlated wtih noise in say the blue channel), again making it
better to do noise removal early on.

2. A fast and effective method is adaptive median filtering
There are a huge range of methods used for noise removal. The
problem is there is always a tradeoff between noise removal and
sharpeness. The adaptive median filter has the interesting
property of being very fast, automatic, and most importantly it
does a good job preserving edges. Essentially the adaptive median
filter takes medians at different angles, then takes the median of
the medians, then if that exceeds a threshold that value is used.

Adaptive median filters don't compete with more advanced techniques
like wavelets or neural networks, but they come close and are nice
and fast.

3. As tools are becoming more advanced, the importance of using raw
data capture is increasing in importance.

In my mind, this makes using the raw DCR format instead of the
processed JPEG-ERI format highliy desirable, as the JPEG-ERI file
has already had the CCD demosaicing applied (and then imagery
converted into the JPEG YCbCr colorspace from RGB). The less
processing done by the camera on the raw file, the better IMHO.

My interest in noise removal came from wanting to clean up longer
time exposure (1 to 60 seconds) photos from the D1x and friends.
I've got results to the point where time exposures look very good.
Noise removal from high ISO rating photos is a harder problem
though than noise removal from time exposure photos.

The results from some of the noise removal programs are very
impressive, especially when considering when they are working from
processed data (say in JPEG format) rather than raw CCD data.



Keyboard shortcuts:
FForum PPrevious NNext WNext unread UUpvote SSubscribe RReply QQuote BBookmark MMy threads
Color scheme? Blue / Yellow