IMI Interdisciplinary Mathematics InstituteCollege of Arts and Sciences

New Frontiers in Imaging and Sensing

headshot Benjamin Berkels
RWTH Aachen

Image Segmentation Based on Learned Discriminative Dictionaries

  • Feb. 22, 2011
  • 9 a.m.
  • Sumwalt 102

Nowadays, sparse signal representations based on overcomplete dictionaries are used for a wide range of signal and image processing tasks. One of the major challenges in this context is the design of suitable dictionaries. The sparse representation itself usually is just a means to an end and used to solve a certain task like, for instance, denoising or compression. Here, we focus on dictionaries suitable for image segmentation tasks and, picking up the discriminative dictionary model by Mairal et al., we introduce an improved minimization algorithm for the underlying variational problem. This algorithm incorporates recent advances in orthogonal matching pursuit made by Rubinstein et al. making it more efficient. Furthermore, it is more stable since it ensures an energy decay in the dictionary update unlike the truncated Newton iteration used by Mairal et al. Finally, we study the applicability of discriminative dictionaries to detect sulci on intra-operative digital photographs of the exposed human cortex. In this application, a discriminative dictionary pair is learned from a set of training images where an experienced physician manually marked the sulci geometry. We demonstrate that this approach allows a robust segmentation of these brain structures as long as the training data contains images sufficiently similar to the input images.

*Joint work with Martin Rumpf (Institute for Numerical Simulation, University of Bonn), Marc Kotowski and Carlo Schaller (University Hospital of Geneva).

© Interdisciplinary Mathematics Institute | The University of South Carolina Board of Trustees | Webmaster