[Insight-users] 2D Registration parameters

[Luis Ibanez]

Hi Nino,

If you know before hand that the images are only translated,
one with respect to the other, you can use the
itkTranslationTranform. This will reduce the parameter space
to just 2D = { xtranslation, ytranslation }.

When you use an AffineTransform in 2D, it provides 6 parameters
because it is formed by a Matrix and a Vector.


A point P = {x,y} is transformed into P' by

                      P' = M * P + V

V is a 2D translation vector, with components { vx, vy }
M is a 2x2 matrix that reprents combinations of rotations, scale
        and shearing.

The parameters {p1,p2,...p6} printed by the example are organized as


M = | p1 p2 |
     | p3 p4 |

V =  | p5 |
      | p6 |


Using an AffineTransform is unnecessary if you know that only
translation is involved in the missalignment. The optimizer is
exploring a 6D parameter space instead of the required 2D
parametric space.

A good combination for 2D images is the one used in

    Insight/Examples/MutualInformationEuler2DRegistration

(today moved to
    Insight/Applications/MutualInformationEuler2DRegistration )

It uses a Euler2D transform
http://www.itk.org/Insight/Doxygen/html/classitk_1_1Euler2DTransform.html

and a MutualInformation metric.
http://www.itk.org/Insight/Doxygen/html/classitk_1_1MutualInformationImageToImageMetric.html
Do your images belong to different modalities.
For example, one taken in an IR band and the other in UV ?
If this is the case, you definetly need to use MutualInformation
as a metric. Otherwise, you may start with a metric like MeanSquares.

A nearest neighbor interpolator should be enough for satelite
images since you usually get very good resolutions.
(is this your case ? I guess that you have at least 1024x1024 pixels...)

The optimizer that is simple to control is the
RegularStepGradientDescentOptimizer
http://www.itk.org/Insight/Doxygen/html/classitk_1_1RegularStepGradientDescentOptimizer.html

If your images are of considerable size, you may want to try the 
multiresolution framework,
http://www.itk.org/Insight/Doxygen/html/classitk_1_1MultiResolutionImageRegistrationMethod.html
whose components are the same used in the basic registration framework.


Please let us know if you have further questions,


   Thanks


    Luis


===========================================
Nino Z wrote:

> Hello Luis,
> 
> I'm trying to register two 2D grayscale satellite images. They are not rotated or scaled. One image is shifted from the other in X and Y direction, so I'm trying to find the X and Y translation values. From the several image registration method tests in the Testing directory, I found several combination of parameters that work with my two test images. When I'm using the Affine transform and the CommandIterator to view the results, I get something like this:
> 
> value: [value1 value2 value3 value4 value5 value6)
> 
>>From the examples I figured out that value5 and value6 are the actual displacement values. In the examples it's [-7 -3]. Are these values in pixels? What do the rest of values mean? Where can I find what the scale and rotation are?
> 
> Is there an optimal combination of registration parameters for the satellite images (optimizer, transform, interpolator and metric) or does the combination depend on the individual images?
> 
> My code is heavily based on the Image Registration example (filtered out all of the FLTK from it) in the Examples directory and I used some of the tricks I learned from the examples in the Algorithms Tests directory.
> 
> Thanks,
> 
> Nino
> 
> 
> 
> 
> _______________________________________________
> Insight-users mailing list
> Insight-users@public.kitware.com
> http://public.kitware.com/mailman/listinfo/insight-users
> 
>