《学习空间上的局部表征》青少年教育丛书.pdfVIP

Learning Spatially Localized, Parts-Based Representation 1 2  1 2 Stan Z. Li , XinWen Hou , HongJiang Zhang , QianSheng Cheng . 1 Microsoft Research China, Beijing Sigma Center, Beijing 100080, China 2 Institute of Mathematical Sciences, Peking University, Beijing, China Contact: szli@, /szli Abstract as object detection and recognition [2, 3, 4, 5, 6, 7]. The significance of is twofold: effective characterization of the In this paper, we propose a novel method, called local non- pattern and dimension reduction. negative matrix factorization (LNMF), for learning spa- One approach for learning a subspace representation for tially localized, parts-based subspace representation of vi- a class of image patterns involves deriving a set of basis sual patterns. An objective function is defined to impose lo- components for construction of the subspace. The eige- calization constraint, in addition to the non-negativity con- niamge method [2, 3, 4] uses principal component analy- straint in the standard NMF [1]. This gives a set of bases sis (PCA) [8] performed on a set of representative training which not only allows a non-subtractive (part-based) repre- data to decorrelate second order moments corresponding to sentation of images but also manifests localized features. low frequency property. Any image can be represented as a An algorithm is presented for the learning of such basis linear combination of these bases. Dimension r