|
| Hi everyone here on comp.sys-soft.matlab!
I've been slowly chugging through an approach I first read about in a
journal article (currently trying to contact the author). I'm feeling
good about the approach, but I have a few issues that I thought you guys
might be able to answer...
I'm trying to quantify a grey-scale image's sharpness. This work is for
electron backscatter patterns obtained in the scanning electron
microscope, incase you were wondering. Since the patterns we obtain are
from diffraction, if the atoms are out of place, or the crystal is
strained, blurriness in the patterns can occur. [This is not the only
way blurriness happens.]
What I'd like to do is compare two very similar images by way of a
quantity of blurriness/sharpness.
So far, what I've come across mentions that a blurry image will have
slow transitions from dark to light in the greyscale, and a sharp image
will have fast transitions from dark to light in the greyscale...
essentially, then, the fft can give us a measure of these frequency
changes... the spectral power density can then give us an amount of
various frequencies contained in the image.
Ok, where I've gotten to- I can calculate the PSD in 2D for these images
(whether these are properly weighted or not is perhaps questionable). I
am taking radial averages from the center of the PSD images, and
producing an average PSD in 1D. This distribution can then give us some
sense of the amount of the frequencies contained in the image.
A few questions-
1) How can I properly compare these average PSD's in 1D to each other?
The proper mean, is the balance point in the distribution and not the
numerical mean, right? Is there a nice way to determine this point?
Contrast will play an important role here if the data is not normalized,
possibly?
2) In the literature they use the first moment of the power spectrum,
however, if I try and take the first moment, I always get zero for the
distribution- if I take the second moment, which I understand is the
variance, my comparisons don't appear to make sense (blurry images don't
appear to be less variant etc.)
I guess I'm having a bit of trouble conceptualizing this. I'll continue
to dig here... but if you have any ideas of how I can compare these
plots numerically (mean/std/etc.) please let me know. Incidentally,
when I very crudely determine the percent of the first peak compared to
the total spectral density value summing over all frequencies in the
image (percent of the first peak area), I do get the trend I was
expecting in terms of ranking for the image blurriness.
Any and all suggestions are greatly appreciated. If you want more
information about my data or about my calculations, please ask!
Thank you for your thoughts!!
-Allen
|
|