| Re: Image registration [message #31194 is a reply to message #25081] |
Wed, 19 June 2002 13:33   |
Eric Vella
Messages: 8
Registered: April 1999
|
Junior Member |
|
|
Just a few additions to the previous good answer from Bill ...
If the X and Y shifts are fairly uncorrelated, and the images have enough
distinctive features (like sharp horizontal and vertical brightness
transitions), you may be able to determine each shift independently. It
might also be possible to proceed iteratively -- determine best rotation,
then best X shift, then best Y shift, then repeat the whole sequence until
it converges. If the convergence is stable, this will be quicker than
mapping out the whole 3D space. If your images are small enough that time
is not a problem, then of course the 3D approach is safer and better.
At each step, you might be able to work with a 1D image projection instead
of the full 2D image. For example, you might be able to determine the X
shift from a projection of the image onto the X axis. For rotation, you
could do something similar, make a 1D vector where each element is a sum of
pixels at a fixed angle, summed over radius.
The C_CORRELATE function lets you evaluate a set of shifts with a single
function call.
"William Thompson" <thompson@orpheus.nascom.nasa.gov> wrote in message
news:aeq5eo$rno$1@skates.gsfc.nasa.gov...
> carsten@rad.uni-kiel.de (Carsten Liess) writes:
>
>> Dear all,
>
>> I am fairly new to IDL, so bear with me please.
>> I have two images of the same object, only with different intensities.
>> I want to create an image that displays the ratio of the two input
>> images. I programmed that and it works.
>> Problem is that the object I am imaging might have moved by up to 10
>> pixels (at a 256*256 matrix) in between, so I need to match the
>> position of these two images before I calculate their ratio. The
>> movement might include translation as well as rotation.
>> Is there a routine in IDL that does such an operation? Alternatively,
>> does any among you have any clever suggestion how to go about this?
>> Many thanks!
>
> Carsten:
>
> The way I've handled this problem in the past is to make a cross
correlation
> matrix of one image against the other, and looking for where the
correlation
> coefficient is highest. Throwing in rotation as well makes it more
> complicated, but in essence the problem is the same. The basic difference
is
> that one is working with three dimensions instead of two.
>
> In the classic problem with translation only, one shifts one image
relative to
> the other by one pixel at a time, and then calculates the cross
correlation
> coefficient between them (e.g. with CORRELATE). For example, if you think
that
> the two images could be shifted relative to each other by a maximum of 10
> pixels, you would build up a matrix of size (21,21) representing shifts
of -10
> to +10 pixels in each direction. If you display this matrix as an image,
you
> should be able to see a clear peak representing the best match between the
two
> images. On my data, I've been able to fit the area around the peak to get
down
> to subpixel resolution--your mileage may vary.
>
> If you add rotation into the mix, I would guess that the best place to
proceed
> would be to build up an array of (N,N,M) of cross correlation
coefficients,
> where N is the number of pixel shifts (21 in the above example), and M is
the
> number of angular rotations you wish to explore. For each angle position
> (K=0,M-1), I would rotate one of the images (e.g. with ROT) and then
calculate
> the NxN cross correlation matrix for that rotation. You would then have a
> three dimensional volume to examine for the peak correlation.
>
> I hope this helps,
>
> Bill Thompson
|
|
|
|
comp.lang.idl-pvwave archive
Members
Search
Help
Login
Home
