On Friday 02 Nov 2012 15:42:57 Thorsten Wagner wrote:
> i've developed a library for extracting the connected components (blobs) of
> a binary ImageJ image. Maybe it's useful for some developers. It is
> freely available on the project homepage:
>
http://code.google.com/p/ijblob/>
> Based on this library, I've developed a plugin for filtering the connected
> components by shape features. What I have to do for publishing it on the
> ImageJ Plugin site?
http://rsb.info.nih.gov/ij/plugins/index.html>
> Kind Regards,
> Thorsten
>
Hi Thorsten,
Thanks for sharing your plugin.
Just a constructive comment about estimating fractal dimensions in that
library:
It is probably tempting, but not a good idea to do it without some control of
the range of boxes that the analysis will use.
In your library, this appears to be hard wired to boxes of up to sides of 64
pixels?
The problems of using arbitrary values like that arise when particle sizes get
close or smaller to that value. That will result in lots of different box
sizes with counts of 1 and the log-log plots will not have the expected linear
trend (hence giving erroneous values of fractal dimension).
Since fractal dimension is a measure of scale invariance, you need a large
range of scales to be sure that such invariance exists. For small particles
this is almost guaranteed to be not so (because their irregularity does not
span lots of scales) and one risks generating numbers that have no meaning.
A simple way of proving this is to repeat the analysis with a slightly
different range of box sizes. If it produces different results, then obviously
there is no scale invariance in the data (and the log-log plots are not
straight).
You will probably read in many publications that one needs at least one order
of magnitude of scales or more for this (so trying to convince others that a
fractal structure exhibits self-similarity over a too-small scale range will
not be easy).
You could also provide for each dimension value estimated some sort of
"goodness of fit" of the fractal dimension. Many people use the R-squared of
the log-log plot (which has its own problems too). Without that, there is no
measure of whether the fractal dimension is a good model to use for the
morphology of that particle (surely Nature has lots of fractal features, but
most certainly not all digitised particles have fractal profiles).
I would suggest caution about estimating fractal dimensions blindly.
I hope my comments are useful.
Cheers
Gabriel
--
ImageJ mailing list:
http://imagej.nih.gov/ij/list.html
Locked
