fiji: registering a pair of images according to another pair of images

[Adrian Daerr-2]

Dear ImageJ users,

I would like to register images taken under very different lighting
conditions. Normal registration fails (tried mainly TurboReg, am trying
others now), I suppose because the colors and intensities change quite a
bit (sometimes the intensity jump across an edge even reverses sign when
light comes from another direction).

If I convert the images to grey images and do an edge detection (e.g.
with FeatureJ's Edges function), these processed images can be very
easily registered. The question is, can I extract the transformation and
apply it to the original pair of RGB images ?

For the moment the closest I get to my goal is to 'Extract SIFT
Correspondences' on the pair of edge-detected images, and transfer the
resulting landmark point selections onto the original images. I now just
lack a registration function which will take these SIFT correspondences
as input. Is something like that present in Fiji ?

I have been looking through the documentation a little bit*, but did not
find an answer yet.

thanks in advance
Adrian


*except for the Fiji main site which looks very strange (letters are all
mixed up when viewed with Firefox or Safari, although the text-only
browser links shows everything correctly). My guess is it's an april
fool's joke.

[Adrian Daerr-2]

> For the moment the closest I get to my goal is to 'Extract SIFT
> Correspondences' on the pair of edge-detected images, and transfer the
> resulting landmark point selections onto the original images. I now just
> lack a registration function which will take these SIFT correspondences
> as input. Is something like that present in Fiji ?

"Moving Least Squares" does the job. Missed that on the first try, sorry
for sending my question a little to early.
Just need to script the whole procedure now.
Still interested to know if there are other means though. I need to do
this for a few hundred photographs, with some masking to exclude truly
changing regions.

cheers,
Adrian

[Stephan Saalfeld]

Hi Adrian,

Plugins->Transform->Landmark Correspondences

http://pacific.mpi-cbg.de/wiki/index.php/Landmark_Correspondences

does the job even better.  You can use the linear (or perspective) model
used for feature extraction or a Moving Least Squares deformation as in
the Moving Least Squares plugin.  Since `Landmark Correspondences' uses
a different renderer and an alternative implementation of the Moving
Least Squares method, it is faster and makes no mistakes.

For batch processing, you might be interested to check another plugin:

Plugins->Registration->Register Virtual Stack Slices

http://pacific.mpi-cbg.de/wiki/index.php/Register_Virtual_Stack_Slices

which can save the transformations as estimated using a variety of
registration methods (feature or intensity based) and

Plugins->Transform->Transform Virtual Stack Slices

http://pacific.mpi-cbg.de/wiki/index.php/Transform_Virtual_Stack_Slices

to apply them to another set of images.

Best,
Stephan




On Fri, 2011-04-01 at 21:54 +0200, Adrian Daerr wrote:

> > For the moment the closest I get to my goal is to 'Extract SIFT
> > Correspondences' on the pair of edge-detected images, and transfer the
> > resulting landmark point selections onto the original images. I now just
> > lack a registration function which will take these SIFT correspondences
> > as input. Is something like that present in Fiji ?
>
> "Moving Least Squares" does the job. Missed that on the first try, sorry
> for sending my question a little to early.
> Just need to script the whole procedure now.
> Still interested to know if there are other means though. I need to do
> this for a few hundred photographs, with some masking to exclude truly
> changing regions.
>
> cheers,
> Adrian