判别字典学习Learning A Discriminative Dictionary for Sparse Representation.pdfVIP

Learning A Discriminative Dictionary for Sparse Representation † ‡ † Zhuolin Jiang , Zhe Lin , Larry S. Davis †Computer Vision Laboratory ‡Advanced Technology Labs University of Maryland, College Park Adobe Systems Incorporated, San Jose {zhuolin, lsd}@ zlin@ 1 Sparse Paradigm  Given a signal y R n and the dictionary D R nK, the sparse representationx R K of y is estimated by: x miny Dx 2  subject to x 0  T F x reconstruction error sparsity constraint 2 Dictionary Learning Techniques  SRC algorithm employs the entire set of training samples to form a dictionary. (face recognition) [Wright, TPAMI2009]  K-SVD: Efficiently learn an over-complete dictionary with a small size. It focuses on representational power, but does not consider discriminative capability. (restoration, compression) [Aharon, TSP2006]  An online algorithm for learning dictionaries is proposed and faster than batch alternatives such as K-SVD on large datasets. (restoration) [Mairal, JMLR 2010]  A tree-structured sparse regularization is proposed to learn a dictionary embedded in a tree efficiently. (restoration) [Jenatton, ICML2010] 3 Dictionary Learning Techniques  Discriminative dictionary learning approaches:  Constructing a separate dictionary for each class [Mairal CVPR2008]  Unifying the dictionary learning and classifier training into a mixed reconstructive and discriminative formulation [Pham,CVPR2008][Zhang,CVPR2010][Mairal,NIPS2009][Yang,CVPR2010]  Solve the problem to alternate between the two variables, minimizing over one while keeping the other one fixed  Learn simultaneously an over-complete dictionary and multiple classi