Loading...

Structured Sparse Representation for Machine Learning and Signal
Processing

Soltani Farani, Ali | 2015

1390 Viewed
  1. Type of Document: Ph.D. Dissertation
  2. Language: Farsi
  3. Document No: 47835 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Rabiei, Hamid Reza
  7. Abstract:
  8. In this proposal we pursue structured sparse representations. Recent years have witnessed a tremendous growth in sparse modeling of natural signals. In this model a signal is represented as a linear combination of a few atoms from an often over-complete dictionary. The recent success of compressed sensing is intact due to the property that natural signals often admit sparse representations. Still, sparse representation in its simplest form sometimes fails to capture the intrinsic structure in natural signals. This structure may be embedded in the signal itself or in the relation between different signals of interest. The goal of this research proposal is to exploit the intrinsic structure in natural signals or the properties of the problem at hand to enhance the effectiveness of sparse representations in machine learning and signal processing applications. To this end we focus on two challenging problems. The first problem is to incorporate the relations among samples into dictionary learning and sparse inference. This relation maybe spatial or temporal. To solve this problem, joint sparse representation and weighted sparse representation are applied to hyperspectral image classification and visual tracking. Joint sparse signals are modeled using the same set of atoms from the dictionary and thus admit the same sparsity pattern. In weighted sparse representation, each dictionary is associated with a specific weight that determines the cost of using that atom in the representation. In this thesis, a group of data collaborate to determine these weights while optimizing the sparse representation. This allows data relations to be incorporated in the model. The second problem is multivariate regression which has many applications in computer vision and image processing. To tackle this problem, the relationship between data and the properties of the problem are quantized in simple assumptions. A novel approach is presented using sparse representation that relies on a common sparse representation in both input and output spaces. The relationship between the data themselves and the dictionary atoms, is used to enforce structure on the sparse representation. This method is applied to 3D human pose estimation. The proposed approaches are evaluated with real data and in comparison to state-ofthe-art methods. A variety of experiments are proposed to assess the properties of the proposed methods
  9. Keywords:
  10. Dictionary Learning ; Sparse Signal Representation ; Structured Sparsity ; Machine Learning Applications

 Digital Object List

 Bookmark

...see more