Tensor-based dictionary learning is a natural approach to form accurate, compressible representations of high-dimensional data (Soltani, Kilmer, Hansen, BIT 2016). In this talk, we explore the generalizability of these tensor dictionaries for image reconstruction problems. For instance, can a dictionary learned from a certain class of images be used to reconstruct a wider variety of images? To reconstruct images efficiently and sparsely from tensor dictionaries, we present a tensor-based Modified Residual Norm Steepest Descent algorithm.

This presentation is part of Contributed Presentation “CP6 - Contributed session 6”

Authors:

Elizabeth Newman (Tufts University)

Misha Kilmer (Tufts University)

Keywords:

image compression, image reconstruction, numerical linear algebra