Hullo Peter,
On Fri, 29 Apr 2011 01:05:59 +1000, Peter Suetterlin
<[email protected]> wrote:
Hi Terry,
haven't been reading the list for some time, but thought I reply anyhow.
I'm a bit puzzled that when using a 512x512 instead of a 256x256 you get
(really?) bad results when you get reasonable ones for the smaller one,
even
if it is contained in the larger one. With the concerns I had about (too
strong) variations of the PSF over the FOV you still should get a
reasonably
reconstructed image in the small subpart that was used to derive it...
I had initially thought that as well, but advice on the subject from a
signal processing group essentially said that the PSF will not be uniform
hence there won't be a good result. I was probably not doing it correctly.
That raises one or two questions on how you are doing things. Does your
PSF
go to 'real' space, or are you doing everything in fourier space?
At that stage of proceedings I had been working only within fourier space.
If you just had two images (sharp and unsharp), FFT them, build the
quotient
and call that PSF, then apply that (Fourier) PSF to the unsharp image
than
indeed this will work perfectly - but only for exactly this image, and
it even
doesn't depend on wether you align it or not (the PSF can/will also
contain
image shift information).
OK.
The other thing is, how did you extend the small PSF to the large FOV?
By
padding with zeros? If this is not done properly things can go bad, too
(FFT
is normally sorted in a wrapped order).
I tried two ways. The first was to split the main image into a number of
subimages equal in size to that from which I derived my PSF, then applied
the PSF to each and finally joined the results together
The second approach was to pad with zeros. That was automagically done by
Octave, so assume it was done properly, also assuming I knew what I was
doing when applying the Octave functions!
One idea what to try would be take the 512 field, divide it into 4x4
fields of
128, derive PSFs in these subfields, average them and apply the average
(propperly padded) to the large image and see what that does. (The idea
is
that if you have only two images to create the PSF it will contain a lot
of
noise information that gets smoothed out that way and make the PSF more
general).
Another one would be to FFT the PSF back to normal space, have a look at
it
and maybe run some filters on it (median to remove singular peaks
etc.). The
abovementioned averaging could also be done on those PSFs...
Thanks for those thoughts/ideas.
For a little while now I have been working on an alternate approach which
attempts to account for non uniform blurring (Whyte et al), and will
continue down this path until I have it working to the point that I can
judge whether it is an approach for this problem, or it defeats me.
I will no doubt come back to where I left off, as the idea of being able
to directly derive a PSF from sharp/blurred pair, seems to be such a good
start to a solution, and try your ideas.
There is a paper (Xianyong Fang, "Feature Based Stitching of a
Clear/Blurred Image
Pair", 2011 International Conference on Multimedia and Signal Processing)
which sounds like it is attacking the same problem, but I haven't been
able to source a copy.
News on my ponderings will probably remain pretty widely spaced for a
while, it is slow progress.
Cheers,
--
Regards,
Terry Duell
--
You received this message because you are subscribed to the Google Groups "Hugin and
other free panoramic software" group.
A list of frequently asked questions is available at:
http://wiki.panotools.org/Hugin_FAQ
To post to this group, send email to [email protected]
To unsubscribe from this group, send email to
[email protected]
For more options, visit this group at http://groups.google.com/group/hugin-ptx
