Modeling the distribution of patches with shift invariance: an application to SAR image restoration.
jeudi 16 octobre 2014, 13h30 - 14h30
Many patch-based denoising procedures aim at learning the distribution of the patches in the image of interest. However they usually do not include shift-invariance, possibly learning redundant patterns. In this presentation we propose a framework to introduce shift invariance by modifying the global optimization function. Instead of searching for a good reconstruction of all the image patches, a good tiling of the image with patches is searched for. A denoising and a dictionary learning schemes with shift invariance are proposed. Besides, in the context of SAR (Synthetic Aperutre Radar) imagery with speckle noise, the fidelity term and associated optimization are adapted.