Sparse Image Representation and its Applications

Day - Time: 22 May 2017, h.15:00
Place: Area della Ricerca CNR di Pisa - Room: C-29
  • Muhammad Hanif (Ercim Fellow at ISTI)

Anna Tonazzini


Recently, sparse representation emerged as a useful regularization in ill-posed linear inverse problems. The key observation in this direction is the fact that many natural signals and images contain highly redundant information and live on low dimension manifolds or subspace that involves much fewer basis compared to the original number of samples. Sparsity is an inherent characteristic of natural signals and can be obtained by decomposing the signal into its elementary basis components, selected either from a predefined basis collection (Wavelets, DCT) or trained over the data (dictionary learning). Images by nature admit a sparse decomposition over a redundant dictionary and sparse representation based methods effectively exploit this property, leading to efficient image processing algorithms. When addressing inverse problem under Bayesian framework with sparsity prior, dictionary is treated as the learning parameter set. Dictionary learning methods, which are central of sparse representation, have been successfully used in a number of signal and image processing applications; i.e. restoration, enhancement, classification, compression, inpainting, super resolution, etc. We will briefly look into some dictionary learning methods and their applications in image processing.