Hullo Kay,
On Mon, 12 Dec 2011 19:42:43 +1100, kfj <[email protected]> wrote:
Terry, I thought some more about the approach to deblur motion-blurred
images by means of 2D-deconvolution. I have doubts the method will
work for wide-angle images. Why is that? Applying a convolution kernel
to a 2D image is comparable to moving the sensor, while the optical
system remains in place. But motion blur comes from moving the whole
optical system together with the sensor. As you go more wide-angle,
the visible effect of the blur should vary from the image center to
the edges. You may be able to deblur sections of the image quite
reliably, but fail to deblur the whole image. So I think you might
have to model the movement of the whole optical system.
I will have to think about this.
I have read a lot of papers on image deblurring and I don't recall any
discussion along these lines. There are approaches that model non-uniform
blur, which is treated as motion in 3 planes, hence a different blur
kernel for each motion plane, but I suspect that is different to what you
are saying.
To model the
blur you could try to superimpose a series of images reprojected to
their original projection after applying incremental amouns of yaw,
pitch and roll to model the blur's underlying movement (and you might
have to add translation, as well). How many images you'd superimpose
in the process would depend on how precisely you want to model the
process. How this approach could be inverted to deblur the image I
can't tell, but maybe one of our resident maths gurus has a word to
say on my wild ideas ;-)
The estimation of a blur kernel using simpler assumptions is a bit of an
exercise. On the surface, your approach might could be order or two more
complex, but you never know.
Just to test my idea, you could try and model a simple blur by
superimposing, say, 10 copies of an image, each yawed by the angle
equivalent to one pixel (so hfov/width) in the image and compare the
result to a 2D blur created by a kernel of ten times .1 in a row. If
you see a noticable difference in the outputs, My argument might be
valid. The superimposition could be done simply by loading the same
image 10 times into hugin, applying the appropropriate yaw to each,
and then do an exposure fusion with settings which assign equal
quality to all images to trick enfuse into simply averaging them, like
choose a big exposure-sigma). What do you say?
Kay
Again, I will have to think about your scheme. I don't think I have any
clear idea of the detailed steps I would use, as you are suggesting use of
hugin in ways that I haven't explored before.
I'm not latching onto 'the angle equiv to one pixel (so hfov/width)'. In
this context what does the 'so' mean?
I'm not sure about your ideas, which doesn't mean they are wrong. Just
that when I am working on a particular approach, the details needed for a
change in direction take a while to percolate...I think it's related to
the 'walk and chew gum' problem :-)
I don't think my knowledge of the solution techniques is good enough to
try to break new ground in this business.
My approach has been to use/adapt the available solution methods/codes for
this problem, hence your approach probably isn't going to help me much as
I won't have the tools to implement it...i.e the kernel estimation will be
the key and that does look a bit daunting.
I have been corresponding with a researcher in the field, and it might be
helpful to ask his advice on your idea, if that's OK.
Thanks for your help.
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
