| Automatic Nonrigid Registration |
| Image Registration and Fusion Systems |
Often the images to be registered have nonlinear geometric differences, requiring a nonlinear transformation function to
register them. A three-step automatic algorithm has been developed by Image Registration and Fusion Systems to register
such images. The algorithm follows the steps below:
method can handle local geometric differences but the differences should vary gradually. The method may not produce accurate
results when images with sharp local geometric differences are used.
Examples of image registration by this software are given below. The images depicted in Figs. 1a and 1b represent two aerial
photographs of the New York City taken in 1966 at scale of 1:24,000. The images have nonlinear geometric differences due to
viewpoint differences. Image geometric differences due to lens nonlinearity and film distortions are also present. These images
are courtesy of NASA. Using the images as input to the automatic nonrigid registration software, first the images are
automatically registered by the affine transformation using three corresponding control points in the images. This result is
shown in Fig. 1c. A dozen corresponding control points are then automatically selected from the registered images. From the
correspondences, the parameters of a surface spline to elastically register the images are obtained. The elastic registration
result is shown in Fig. 1d. The user may now require the software to choose a larger number of correspondences to improve
the registration accuracy. By the press of a mouse button, the program automatically selects a larger number of control point
correspondences in the images and registers the images. Fig. 1e shows registration using 39 control point correspondences.
A larger number of correspondences does not necessarily mean a more accurate registration. As more control points are
selected, the likelihood of obtaining inaccurate correspondences increases, and that could reduce the overall registration
accuracy. The program, however, can detect this and stop the process. The program works ideally when the images do not
have intensity differences. The process that stops the registration will become less accurate as intensity differences between
the images increase.
A second example is given in Fig. 2. The images to be registered are given in Figs. 2a and 2b and the result after the affine
registration is shown in Fig. 2c. Ten control point correspondences are used to find the affine parameters. The number of
control points selected in each step is image dependent. When more than three control points are found, a process selects the
best three to maximize accuracy in affine registration. The result of nonrigid registration using 19 control points is shown in Fig.
2d. Again, the number of control points selected and used in nonrigid registration is image dependent. Asking the program to
select and use more control points in these images actually reduces the registration accuracy due to the rather sharp local
geometric differences between the images, Therefore, the process stops after initial nonrigid registration.
This software is suitable for registration of aerial images. The software is also suitable for registration of biomedical images
where intensity differences between the images are small, occlusion is rare in the images, and local geometric differences
between images are very small.
For more information about this automatic nonrigid registration software contact Image Registration and Fusion Systems.
register them. A three-step automatic algorithm has been developed by Image Registration and Fusion Systems to register
such images. The algorithm follows the steps below:
- Select a small number of corresponding landmarks in the images and from the correspondences determine the
parameters of an affine transformation to approximately registers the images. - Select a larger number of landmark correspondences in the approximately registered images and from the
correspondences determine the parameters of a surface spline to register the images. - Repeat step 2 as needed to increase the registration accuracy.
method can handle local geometric differences but the differences should vary gradually. The method may not produce accurate
results when images with sharp local geometric differences are used.
Examples of image registration by this software are given below. The images depicted in Figs. 1a and 1b represent two aerial
photographs of the New York City taken in 1966 at scale of 1:24,000. The images have nonlinear geometric differences due to
viewpoint differences. Image geometric differences due to lens nonlinearity and film distortions are also present. These images
are courtesy of NASA. Using the images as input to the automatic nonrigid registration software, first the images are
automatically registered by the affine transformation using three corresponding control points in the images. This result is
shown in Fig. 1c. A dozen corresponding control points are then automatically selected from the registered images. From the
correspondences, the parameters of a surface spline to elastically register the images are obtained. The elastic registration
result is shown in Fig. 1d. The user may now require the software to choose a larger number of correspondences to improve
the registration accuracy. By the press of a mouse button, the program automatically selects a larger number of control point
correspondences in the images and registers the images. Fig. 1e shows registration using 39 control point correspondences.
A larger number of correspondences does not necessarily mean a more accurate registration. As more control points are
selected, the likelihood of obtaining inaccurate correspondences increases, and that could reduce the overall registration
accuracy. The program, however, can detect this and stop the process. The program works ideally when the images do not
have intensity differences. The process that stops the registration will become less accurate as intensity differences between
the images increase.
A second example is given in Fig. 2. The images to be registered are given in Figs. 2a and 2b and the result after the affine
registration is shown in Fig. 2c. Ten control point correspondences are used to find the affine parameters. The number of
control points selected in each step is image dependent. When more than three control points are found, a process selects the
best three to maximize accuracy in affine registration. The result of nonrigid registration using 19 control points is shown in Fig.
2d. Again, the number of control points selected and used in nonrigid registration is image dependent. Asking the program to
select and use more control points in these images actually reduces the registration accuracy due to the rather sharp local
geometric differences between the images, Therefore, the process stops after initial nonrigid registration.
This software is suitable for registration of aerial images. The software is also suitable for registration of biomedical images
where intensity differences between the images are small, occlusion is rare in the images, and local geometric differences
between images are very small.
For more information about this automatic nonrigid registration software contact Image Registration and Fusion Systems.
(a)
(b)
(c)
(d)
(e)
Fig. 1. (a), (b) Aerial photographs of the New York City taken in 1966. These images are courtesy of NASA. Image (b) has been
rotated by 90 degrees from its original orientation to increase the global geometric difference between the images. (c) The
registration result after the affine registration. (d) The result after the initial nonrigid registration. (e) The result after repeating
the nonrigid registration one more time using a larger number of control point correspondences.
rotated by 90 degrees from its original orientation to increase the global geometric difference between the images. (c) The
registration result after the affine registration. (d) The result after the initial nonrigid registration. (e) The result after repeating
the nonrigid registration one more time using a larger number of control point correspondences.
(a)
(b)
(c)
(d)
Fig. 2. (a), (b) Two images of an outdoor scene taken from slightly different views. (c) Affine registration result. (d) Nonrigid
registration result.
registration result.
 | |  |
To obtain a license for this automatic nonrigid image registration software, follow this link =>