Metamorphosis of Images in Reproducing Kernel Hilbert Spaces › ~dlabate › Zhao_talk.pdfIn...
Transcript of Metamorphosis of Images in Reproducing Kernel Hilbert Spaces › ~dlabate › Zhao_talk.pdfIn...
![Page 1: Metamorphosis of Images in Reproducing Kernel Hilbert Spaces › ~dlabate › Zhao_talk.pdfIn applications, we always need to compare the similarity of two di erent shapes; di eomorphic](https://reader035.fdocuments.us/reader035/viewer/2022063001/5f1b979f56d5753e5952ee9f/html5/thumbnails/1.jpg)
Metamorphosis of Images in Reproducing KernelHilbert Spaces
by Heng Zhao
May 6, 2020
by Heng Zhao Metamorphosis of Images in Reproducing Kernel Hilbert Spaces May 6, 2020 1 / 25
![Page 2: Metamorphosis of Images in Reproducing Kernel Hilbert Spaces › ~dlabate › Zhao_talk.pdfIn applications, we always need to compare the similarity of two di erent shapes; di eomorphic](https://reader035.fdocuments.us/reader035/viewer/2022063001/5f1b979f56d5753e5952ee9f/html5/thumbnails/2.jpg)
Content
What is diffeomorphic matching of shapes?
How to build diffeomorphisms
Large Deformation Diffeomorphic Metric Mapping.
Shooting Method.
Numerical Experiments.
Results on MNIST
Conclusion
by Heng Zhao Metamorphosis of Images in Reproducing Kernel Hilbert Spaces May 6, 2020 2 / 25
![Page 3: Metamorphosis of Images in Reproducing Kernel Hilbert Spaces › ~dlabate › Zhao_talk.pdfIn applications, we always need to compare the similarity of two di erent shapes; di eomorphic](https://reader035.fdocuments.us/reader035/viewer/2022063001/5f1b979f56d5753e5952ee9f/html5/thumbnails/3.jpg)
What is diffeomorphic matching of shapes?
Metamorphosis is a method for diffeomorphic matching of shapes, with many potential applications for anatomical shape comparison in medical image analysis; it is a central problem in the field of computational anatomy.
In applications, we always need to compare the similarity of twodifferent shapes; diffeomorphic matching will provide a way to compare shapes. Two shapes are similar if one can be obtained from the other via a small deformation.
by Heng Zhao Metamorphosis of Images in Reproducing Kernel Hilbert Spaces May 6, 2020 3 / 25
![Page 4: Metamorphosis of Images in Reproducing Kernel Hilbert Spaces › ~dlabate › Zhao_talk.pdfIn applications, we always need to compare the similarity of two di erent shapes; di eomorphic](https://reader035.fdocuments.us/reader035/viewer/2022063001/5f1b979f56d5753e5952ee9f/html5/thumbnails/4.jpg)
How to build diffeomorphisms
For t ∈ [0, 1], velocity field v(t) : Rd → Rd . The position x(t) ∈ R attime t of a particle that moves along this velocity field is described by
dx
dt(x) = v (t, x (t)) . (1)
This is a deformation of the space at time t, denoted ϕ (t), soϕ (t, x) is the position of particle at time t started its motion at x attime 0. In particular, ϕ (0, x) = x .
As long as we take v(t) ”very regular” with respect to the spacevariables, the transformation will be a diffeomorphism: it will mapsmooth curves onto smooth curves, corners onto corners, and preservepresence or lack of self-intersection points.
by Heng Zhao Metamorphosis of Images in Reproducing Kernel Hilbert Spaces May 6, 2020 4 / 25
![Page 5: Metamorphosis of Images in Reproducing Kernel Hilbert Spaces › ~dlabate › Zhao_talk.pdfIn applications, we always need to compare the similarity of two di erent shapes; di eomorphic](https://reader035.fdocuments.us/reader035/viewer/2022063001/5f1b979f56d5753e5952ee9f/html5/thumbnails/5.jpg)
How to build diffeomorphisms
Figure: Initial Grid.
by Heng Zhao Metamorphosis of Images in Reproducing Kernel Hilbert Spaces May 6, 2020 5 / 25
![Page 6: Metamorphosis of Images in Reproducing Kernel Hilbert Spaces › ~dlabate › Zhao_talk.pdfIn applications, we always need to compare the similarity of two di erent shapes; di eomorphic](https://reader035.fdocuments.us/reader035/viewer/2022063001/5f1b979f56d5753e5952ee9f/html5/thumbnails/6.jpg)
How to build diffeomorphisms
Figure: A controller specifies a direction at every point, at every time.
by Heng Zhao Metamorphosis of Images in Reproducing Kernel Hilbert Spaces May 6, 2020 6 / 25
![Page 7: Metamorphosis of Images in Reproducing Kernel Hilbert Spaces › ~dlabate › Zhao_talk.pdfIn applications, we always need to compare the similarity of two di erent shapes; di eomorphic](https://reader035.fdocuments.us/reader035/viewer/2022063001/5f1b979f56d5753e5952ee9f/html5/thumbnails/7.jpg)
How to build diffeomorphisms
Figure: A controller specifies a direction at every point, at every time.
by Heng Zhao Metamorphosis of Images in Reproducing Kernel Hilbert Spaces May 6, 2020 7 / 25
![Page 8: Metamorphosis of Images in Reproducing Kernel Hilbert Spaces › ~dlabate › Zhao_talk.pdfIn applications, we always need to compare the similarity of two di erent shapes; di eomorphic](https://reader035.fdocuments.us/reader035/viewer/2022063001/5f1b979f56d5753e5952ee9f/html5/thumbnails/8.jpg)
How to build diffeomorphisms
Figure: A controller specifies a direction at every point, at every time.
by Heng Zhao Metamorphosis of Images in Reproducing Kernel Hilbert Spaces May 6, 2020 8 / 25
![Page 9: Metamorphosis of Images in Reproducing Kernel Hilbert Spaces › ~dlabate › Zhao_talk.pdfIn applications, we always need to compare the similarity of two di erent shapes; di eomorphic](https://reader035.fdocuments.us/reader035/viewer/2022063001/5f1b979f56d5753e5952ee9f/html5/thumbnails/9.jpg)
How to build diffeomorphisms
Figure: A controller specifies a direction at every point, at every time.
by Heng Zhao Metamorphosis of Images in Reproducing Kernel Hilbert Spaces May 6, 2020 9 / 25
![Page 10: Metamorphosis of Images in Reproducing Kernel Hilbert Spaces › ~dlabate › Zhao_talk.pdfIn applications, we always need to compare the similarity of two di erent shapes; di eomorphic](https://reader035.fdocuments.us/reader035/viewer/2022063001/5f1b979f56d5753e5952ee9f/html5/thumbnails/10.jpg)
How to build diffeomorphisms
Figure: A controller specifies a direction at every point, at every time.
by Heng Zhao Metamorphosis of Images in Reproducing Kernel Hilbert Spaces May 6, 2020 10 / 25
![Page 11: Metamorphosis of Images in Reproducing Kernel Hilbert Spaces › ~dlabate › Zhao_talk.pdfIn applications, we always need to compare the similarity of two di erent shapes; di eomorphic](https://reader035.fdocuments.us/reader035/viewer/2022063001/5f1b979f56d5753e5952ee9f/html5/thumbnails/11.jpg)
How to build diffeomorphisms
Figure: A controller specifies a direction at every point, at every time.
by Heng Zhao Metamorphosis of Images in Reproducing Kernel Hilbert Spaces May 6, 2020 11 / 25
![Page 12: Metamorphosis of Images in Reproducing Kernel Hilbert Spaces › ~dlabate › Zhao_talk.pdfIn applications, we always need to compare the similarity of two di erent shapes; di eomorphic](https://reader035.fdocuments.us/reader035/viewer/2022063001/5f1b979f56d5753e5952ee9f/html5/thumbnails/12.jpg)
How to build diffeomorphisms
Figure: A controller specifies a direction at every point, at every time.
by Heng Zhao Metamorphosis of Images in Reproducing Kernel Hilbert Spaces May 6, 2020 12 / 25
![Page 13: Metamorphosis of Images in Reproducing Kernel Hilbert Spaces › ~dlabate › Zhao_talk.pdfIn applications, we always need to compare the similarity of two di erent shapes; di eomorphic](https://reader035.fdocuments.us/reader035/viewer/2022063001/5f1b979f56d5753e5952ee9f/html5/thumbnails/13.jpg)
How to build diffeomorphisms
Figure: A controller specifies a direction at every point, at every time.
by Heng Zhao Metamorphosis of Images in Reproducing Kernel Hilbert Spaces May 6, 2020 13 / 25
![Page 14: Metamorphosis of Images in Reproducing Kernel Hilbert Spaces › ~dlabate › Zhao_talk.pdfIn applications, we always need to compare the similarity of two di erent shapes; di eomorphic](https://reader035.fdocuments.us/reader035/viewer/2022063001/5f1b979f56d5753e5952ee9f/html5/thumbnails/14.jpg)
How to build diffeomorphisms
Figure: A controller specifies a direction at every point, at every time.
by Heng Zhao Metamorphosis of Images in Reproducing Kernel Hilbert Spaces May 6, 2020 14 / 25
![Page 15: Metamorphosis of Images in Reproducing Kernel Hilbert Spaces › ~dlabate › Zhao_talk.pdfIn applications, we always need to compare the similarity of two di erent shapes; di eomorphic](https://reader035.fdocuments.us/reader035/viewer/2022063001/5f1b979f56d5753e5952ee9f/html5/thumbnails/15.jpg)
How to build diffeomorphisms
Figure: A controller specifies a direction at every point, at every time.
by Heng Zhao Metamorphosis of Images in Reproducing Kernel Hilbert Spaces May 6, 2020 15 / 25
![Page 16: Metamorphosis of Images in Reproducing Kernel Hilbert Spaces › ~dlabate › Zhao_talk.pdfIn applications, we always need to compare the similarity of two di erent shapes; di eomorphic](https://reader035.fdocuments.us/reader035/viewer/2022063001/5f1b979f56d5753e5952ee9f/html5/thumbnails/16.jpg)
How to build diffeomorphisms
Figure: Final Grid.
by Heng Zhao Metamorphosis of Images in Reproducing Kernel Hilbert Spaces May 6, 2020 16 / 25
![Page 17: Metamorphosis of Images in Reproducing Kernel Hilbert Spaces › ~dlabate › Zhao_talk.pdfIn applications, we always need to compare the similarity of two di erent shapes; di eomorphic](https://reader035.fdocuments.us/reader035/viewer/2022063001/5f1b979f56d5753e5952ee9f/html5/thumbnails/17.jpg)
Large Deformation Diffeomorphic Metric Mapping
Fix a shape q0, the template, from which we want to register anothershape q1 (the target).
A time-dependent velocity field (t, x)→ v(t, x) yields a deformation(t, x)→ ϕ(t, x), which acts onto q0 as denoted by q(t) := ϕ(t) · q0.The goal is now to find v∗ which minimizes a functional
J(v) =1
2
∫ 1
0‖v(t)‖2V dt + U(q(1))→ min (2)
subject to ∂t q(t) = v(t) · q(t), q(0) = qwhere ‖·‖V is an appropriate Hilbert norm. The data attachment U(q(1)) is a crude measure of the difference between the deformed shape q(1) and the target q1.
by Heng Zhao Metamorphosis of Images in Reproducing Kernel Hilbert Spaces May 6, 2020 17 / 25
![Page 18: Metamorphosis of Images in Reproducing Kernel Hilbert Spaces › ~dlabate › Zhao_talk.pdfIn applications, we always need to compare the similarity of two di erent shapes; di eomorphic](https://reader035.fdocuments.us/reader035/viewer/2022063001/5f1b979f56d5753e5952ee9f/html5/thumbnails/18.jpg)
Large Deformation Diffeomorphic Metric Mapping
In order to connect two images q(0) and q(1) in H with a continuouspath q(t), image metamorphosis solves the optimal controlproblem(3):
by Heng Zhao Metamorphosis of Images in Reproducing Kernel Hilbert Spaces May 6, 2020 18 / 25
![Page 19: Metamorphosis of Images in Reproducing Kernel Hilbert Spaces › ~dlabate › Zhao_talk.pdfIn applications, we always need to compare the similarity of two di erent shapes; di eomorphic](https://reader035.fdocuments.us/reader035/viewer/2022063001/5f1b979f56d5753e5952ee9f/html5/thumbnails/19.jpg)
Large Deformation Diffeomorphic Metric Mapping
Well known results on RKHS imply that these optimal solutions must be of the form
by Heng Zhao Metamorphosis of Images in Reproducing Kernel Hilbert Spaces May 6, 2020 19 / 25
![Page 20: Metamorphosis of Images in Reproducing Kernel Hilbert Spaces › ~dlabate › Zhao_talk.pdfIn applications, we always need to compare the similarity of two di erent shapes; di eomorphic](https://reader035.fdocuments.us/reader035/viewer/2022063001/5f1b979f56d5753e5952ee9f/html5/thumbnails/20.jpg)
Large Deformation Diffeomorphic Metric Mapping
for some coefficients z and α, and that their norms are given by
by Heng Zhao Metamorphosis of Images in Reproducing Kernel Hilbert Spaces May 6, 2020 20 / 25
![Page 21: Metamorphosis of Images in Reproducing Kernel Hilbert Spaces › ~dlabate › Zhao_talk.pdfIn applications, we always need to compare the similarity of two di erent shapes; di eomorphic](https://reader035.fdocuments.us/reader035/viewer/2022063001/5f1b979f56d5753e5952ee9f/html5/thumbnails/21.jpg)
Large Deformation Diffeomorphic Metric Mapping
Solutions of (3) are therefore solutions of the reduced problem (4):
by Heng Zhao Metamorphosis of Images in Reproducing Kernel Hilbert Spaces May 6, 2020 21 / 25
![Page 22: Metamorphosis of Images in Reproducing Kernel Hilbert Spaces › ~dlabate › Zhao_talk.pdfIn applications, we always need to compare the similarity of two di erent shapes; di eomorphic](https://reader035.fdocuments.us/reader035/viewer/2022063001/5f1b979f56d5753e5952ee9f/html5/thumbnails/22.jpg)
Shooting Method
Then this paper derives the shooting method for (4)
by Heng Zhao Metamorphosis of Images in Reproducing Kernel Hilbert Spaces May 6, 2020 22 / 25
![Page 23: Metamorphosis of Images in Reproducing Kernel Hilbert Spaces › ~dlabate › Zhao_talk.pdfIn applications, we always need to compare the similarity of two di erent shapes; di eomorphic](https://reader035.fdocuments.us/reader035/viewer/2022063001/5f1b979f56d5753e5952ee9f/html5/thumbnails/23.jpg)
Numerical Experiments
For numerical experiment, they use
with u = |x − y | /τV and u˜ = |x − y | /τH , where τV and τH arewidth parameters associated to the reproducing kernels1.
In this numerical experiment, they use τV = 1.5 and τH = 0.5
by Heng Zhao Metamorphosis of Images in Reproducing Kernel Hilbert Spaces May 6, 2020 23 / 25
![Page 24: Metamorphosis of Images in Reproducing Kernel Hilbert Spaces › ~dlabate › Zhao_talk.pdfIn applications, we always need to compare the similarity of two di erent shapes; di eomorphic](https://reader035.fdocuments.us/reader035/viewer/2022063001/5f1b979f56d5753e5952ee9f/html5/thumbnails/24.jpg)
Results
Figure: Morphing of letter D from MNIST training set: top row shows evolution of the template; bottom row shows evolution of deformed template.
by Heng Zhao Metamorphosis of Images in Reproducing Kernel Hilbert Spaces May 6, 2020 24 / 25
![Page 25: Metamorphosis of Images in Reproducing Kernel Hilbert Spaces › ~dlabate › Zhao_talk.pdfIn applications, we always need to compare the similarity of two di erent shapes; di eomorphic](https://reader035.fdocuments.us/reader035/viewer/2022063001/5f1b979f56d5753e5952ee9f/html5/thumbnails/25.jpg)
Conclusion
In this paper, they proposed a particle based optimization method for their estimation and a shooting method based on a specific reproducing kernels. This algorithm allows for numerically stable sparse representation of the target image in a template-centered coordinate system, which is hard to obtain using former methods. The introduction of the Sobolev Space norm for the images plays a critical role as it allows for particle solutions that would not be possible using an L2 norm.
by Heng Zhao Metamorphosis of Images in Reproducing Kernel Hilbert Spaces May 6, 2020 25 / 25