Numerical aspects of spectral segmentation on polygonal grids
- Los Alamos National Laboratory
- UNIV OF CAMBRIDGE
The authors analyze numerical behavior of the spectral graph partitioning problem arising in the Normalized Cuts formulation of the image segmentation problem on polygonal grids. They make an observation that in the presence of rounding errors the eigenvector corresponding to the k-th smallest eigenvalue of the generalized graph Laplacian should contain more than k nodal domains that represent coherent segments in the image. As the result, the eigenvector corresponding to the trivial solution carries a wealth of information about the nodal domains in the image and can be used as an initial guess for the Krylov subspace eigensolver, while the computed eigenvector subspace, corresponding to just a few of the lowest eigenvalues of the graph Laplacian, will contain sufficient information for obtaining meaningful segmentation.
- Research Organization:
- Los Alamos National Laboratory (LANL), Los Alamos, NM (United States)
- Sponsoring Organization:
- USDOE
- DOE Contract Number:
- AC52-06NA25396
- OSTI ID:
- 1000927
- Report Number(s):
- LA-UR-10-01628; LA-UR-10-1628; TRN: US201101%%784
- Resource Relation:
- Journal Volume: 7133; Conference: Para 2010: State of the Art in Scientific and Parallel Computing ; June 6, 2010 ; Reykjavik, Iceland
- Country of Publication:
- United States
- Language:
- English
Similar Records
Analysis of Krylov subspace recycling for sequences of linear systems.
Patch-based image segmentation of satellite imagery using minimum spanning tree construction