In this work, we develop a novel solution to the problem of eciently encoding multiresolution pyramids of big data images. The novelty of our approach lies in the transformation of a regular image into a probability density function image in the space range domain. We use the popular convolutional sparse coding (csc) method to encode images and formulate an `1 constrained optimization problem to enforce sparsity on the csc- coecients. This problem is particularly challenging because in general it is non-convex. We use the alternating direction method of multipliers (admm) based on the augmented Lagrange function to solve it iteratively by splitting the global objective into a sum of convex subproblems. To enhance performance, we carry out convolution in the spectral domain. To enhance convergence we work with the proximal operators framework. Our results yield good approximation quality with all image types and demonstrate, that the method's `1 regularization enforces sparsity in all coecient maps.
Masterprüfung mit Defensio, Wozniak Kinga Anna
25.09.2018 13:30 - 15:00
Organiser:
SPL 5
Location:
Besprechungsraum 4.34
Währinger Straße 29
1090 Wien
1090 Wien