 |
Center for Nonlinear Studies |
Sparse representations with learned dictionaries are one of the leading image modeling techniques for image restoration. When learning these dictionaries from a set of training images, the sparsity parameter of the dictionary learning algorithm strongly influences the content of the dictionary atoms, and ultimately the performance of an image restoration technique making use of that dictionary. In this talk, we describe geometrically the content of trained dictionaries and how it changes with the sparsity parameter. We use statistical analysis to characterize how the different content is used in sparse representations. Finally, we demonstrate a method to control the structure of the dictionaries, allowing us to learn a dictionary which can later be tailored for specific applications.
Host: Brendt Wohlberg