MA: Low rank- and sparsity-based image registration

Image Registration Experiments

First simple experiments for non-parametric image registration, written in MATLAB. Provides:

• Distance Measures

  • SSD

• Regularizers

  • Diffusive Energy
  • Curvature Energy

• Optimization schemes

  • Gradient descent
  • Gauß-Newton optimization
  • Armijo line search
  • Support for Multi-Level strategy

• Miscellaneous

  • Derivative test (1st + 2nd order) for multivariate functions

Primal Dual Optimization

Convex optimization experiments with first-order primal-dual algorithm by Chambolle & Pock. Provides:

• TV-L1 Image Denoising

• TV-L2 and TV-L1 Image Registration

Note: Image Registration procedures use an iterative linear approximation of the image model to achieve a convex data term. Details can be found in A Duality Based Algorithm for TV-L1-Optical-Flow Image Registration.

Nuclear Norm Experiments

• A new distance measure for simultaneous image registration of an arbitrary number of template images (omitting a predefined reference). The rough idea is to constrain the rank of the matrix of column-major images, thus enforcing similarity between the images. Based on (and modified from) Shape from Light Field Meets Robust PCA.

• Optimization is performed in a similiar fashion as the TV-L1 and TV-L2 registration from above, i.e. using convex image model approximations and applying the primal-dual algorithm by Chambolle & Pock.

This project and all code included with it is licensed under the MIT Open Source Lincense (see LICENSE file for details). If you have questions, contact me at roland.haase [at] student.uni-luebeck [dot] de.