Image Registration with Homography: A Refresher with Differentiable Mutual Information, Ordinary Differential Equation and Complex Matrix Exponential

  • Author / Creator
    Nan, Abhishek
  • This work presents a novel method of tackling the task of image registration. Our algorithm uses a differentiable form of Mutual Information implemented via a neural network called MINE. An important property of neural networks is them being differentiable, which allows them to be used as a loss function. This way we use MINE as an estimator for our loss function. Furthermore to make the optimization smoother, we parametrize the transformation module using complex matrix manifolds which further improves our accuracy and efficiency. In order to speed up computation and make the algorithm more robust we use a multi-resolution approach, but implement it as a simultaneous loss from all levels, which provides the aforementioned benefits. The parameters for each resolution are modelled via ordinary differential equations and solved using a neural network which adds to the final performance scores as well. This leads to a state of the art algorithm implemented via modern software frameworks which allow for automatic gradient computations (such as PyTorch). Our algorithm performs better than registration algorithms available off the shelf in state of the art image registration tools/softwares. We demonstrate this on four open source datasets.

  • Subjects / Keywords
  • Graduation date
    Fall 2020
  • Type of Item
  • Degree
    Master of Science
  • DOI
  • License
    Permission is hereby granted to the University of Alberta Libraries to reproduce single copies of this thesis and to lend or sell such copies for private, scholarly or scientific research purposes only. Where the thesis is converted to, or otherwise made available in digital form, the University of Alberta will advise potential users of the thesis of these terms. The author reserves all other publication and other rights in association with the copyright in the thesis and, except as herein before provided, neither the thesis nor any substantial portion thereof may be printed or otherwise reproduced in any material form whatsoever without the author's prior written permission.