Posted by
Frederick Ross on
Apr 25, 2007; 1:59pm
URL: http://imagej.273.s1.nabble.com/help-standardizing-images-tp3699688p3699690.html
As Gabriel said, please post a picture of your card.
On 4/24/07, Ryan Garrick <
[hidden email]> wrote:
> Frederick, it looks like the procedure you suggested has worked quite
> well. I haven't yet played around with equalization (or stretching, as
> mentioned by Gabriel) that much, but might need to because although the
> graycard in my reference image has a histogram of gray values with an
> approx. normal distribution, the graycards in some of target images give
> a histogram with a slightly longer upper tail (c.f. a normal
> distribution). If I use the min. and max. gray values alone to determine
> scaling and offset, this tail is likely to have some impact on
> standardization. Perhaps it would be reasonable to use the min. and modal
> gray value of the histograms, the rationale being that it is more
> important end up with comparable frequency distributions of gray values
> rather than matching range limits?
If you get distortions of the original Gaussian, you need to NOT use
the procedure I recommended, as it will introduce systematic errors
into your numbers. It means that your sensor is not linear (as
Gabriel had mentioned before). The best thing is to keep your sensor
in a range where it's linear, but I realize that's not always
possible.
This being so, life is rather harder. Essentially, you've got to fit
a function to the histogram for each of two images, then make the
coordinate transformation that takes one to another. For this reason
you'll want fairly manageable functions. You might try G(x^\alpha)
where G is the Gaussian, and alpha is some power close to 1 (so you're
fitting three parameters: the mean, standard deviation, and alpha).
If your sensor isn't too far from linear, you may be able to get
satisfactory fits this way, and the transformation is fairly
straightforwards.
I personally think the easiest way is to call Octave from a plugin
(JAPOS at
http://jopas.sourceforge.net/ lets you do this reasonably
well), and then run some curve fitting algorithm someone's already
written there.
Otherwise, normally for a Gaussian you are better off measuring the
mean and standard deviation. These will be more robust than min and
max.
--
Frederick Ross
Graduate Fellow, (|Siggia> + |McKinney>)/sqrt(2) Lab
The Rockefeller University
Je ne suis pas Fred Cross!