Point Coding: Sparse Image Representation with Adaptive Shiftable-Kernel Dictionaries
Résumé
This paper addresses the problem of adaptively deriving optimally sparse image representations, using an dictionary composed of shiftable kernels. Algorithmic advantages of our solution make possible the computation of an approximately shift-invariant adaptive image representation. Learned kernels can have different sizes and adapt to different scales. Coefficient extraction uses a fast implementation of Matching Pursuit with essentially logarithmic cost per iteration. Dictionary update is performed by solving a structured least-squares problem either by algebraic characterization of pseudoinverses of structured matrices, or by superfast interpolation methods. Kernels learned from natural images display expected 2D Gabor aspect (localization in orientation and frequency), as well as other structures commonly occurring in images (e.g., curved edges, or cross patterns), while when applied to newspaper text images, kernels tend to reproduce printed symbols or groups thereof.
Origine : Fichiers produits par l'(les) auteur(s)
Loading...