|
Image Modeling and Enhancement with Structured Sparse Mixing Estimators| old_uid | 8436 |
|---|
| title | Image Modeling and Enhancement with Structured Sparse Mixing Estimators |
|---|
| start_date | 2010/03/25 |
|---|
| schedule | 12h-13h |
|---|
| online | no |
|---|
| summary | Accurate signal estimations for denoising and inverse problems may be recovered by selecting a sparse set of vectors in a redundant dictionary. Reviewing this generic approach shows that a complete freedom of choice comes with a high computational complexity, and it often increases the estimation risk. Structured sparse approximations are introduced in dictionaries which are learned from the input data as a union of linear orthogonal bases, using Gaussian mixture models. A linear estimator is computed in each orthogonal basis and the best estimator is selected for signal estimation. The degree of freedom is significantly lower than with traditional overcomplete dictionary approaches, stabilizing the estimation. The computational complexity is typically 2 orders of magnitude faster than estimation in an overcomplete dictionary. State-of-the-art results are shown in image denoising, inpainting, super-resolution interpolation and deblurring. |
|---|
| responsibles | Aujol, Beauchard, Pauchont |
|---|
| |
|