Robust Alternating Low-Rank Representation by joint L p - and L 2,p -norm minimization.

Autor: Zhang Z; School of Computer Science and Technology & Joint International Research Laboratory of Machine Learning and Neuromorphic Computing, Soochow University, Suzhou 215006, China. Electronic address: cszzhang@gmail.com., Zhao M; Department of Electronic Engineering, City University of Hong Kong, Tat Chee Avenue, Kowloon, Hong Kong., Li F; School of Computer Science and Technology & Joint International Research Laboratory of Machine Learning and Neuromorphic Computing, Soochow University, Suzhou 215006, China., Zhang L; School of Computer Science and Technology & Joint International Research Laboratory of Machine Learning and Neuromorphic Computing, Soochow University, Suzhou 215006, China., Yan S; Department of Electrical and Computer Engineering, National University of Singapore, Singapore.
Jazyk: angličtina
Zdroj: Neural networks : the official journal of the International Neural Network Society [Neural Netw] 2017 Dec; Vol. 96, pp. 55-70. Date of Electronic Publication: 2017 Sep 14.
DOI: 10.1016/j.neunet.2017.08.001
Abstrakt: We propose a robust Alternating Low-Rank Representation (ALRR) model formed by an alternating forward-backward representation process. For forward representation, ALRR first recovers the low-rank PCs and random corruptions by an adaptive local Robust PCA (RPCA). Then, ALRR performs a joint L p -norm and L 2,p -norm minimization (0

p -norm on the coefficients can ensure joint sparsity for subspace representation, while the L 2,p -norm on the reconstruction error can handle outlier pursuit. After that, ALRR returns the coefficients as adaptive weights to local RPCA for updating PCs and dictionary in the backward representation process. Thus, ALRR is regarded as an integration of local RPCA with adaptive weights plus sparse LRR with a self-expressive low-rank dictionary. To enable ALRR to handle outside data efficiently, a projective ALRR that can extract features from data directly by embedding is also derived. To solve the L 2,p -norm based minimization problem, a new iterative scheme based on the Iterative Shrinkage/Thresholding (IST) approach is presented. The relationship analysis with other related criteria show that our methods are more general. Visual and numerical results demonstrate the effectiveness of our algorithms for representation.
(Copyright © 2017 Elsevier Ltd. All rights reserved.)
Databáze: MEDLINE
Externí odkaz: