期刊名称：International Journal of Mechatronics, Electrical and Computer Technology
印刷版ISSN：2305-0543
出版年度：2016
卷号：6
期号：20
页码：2878-2889
出版社：Austrian E-Journals of Universal Scientific Organization
摘要：NMF algorithms have been recently employed in several applications; however, the performance of NMF is highly dependent on three factors including: 1) choosing a suitable cost function, 2) using an effective initialization method and 3) determining the rank of factorized matrices. This paper is aimed at enhancing the NMF performance using CRF as an efficient initialization method for estimating initial components of NMF in addition to find the proper rank of decomposed matrices. The modified NMF is applied to JAFFE facial expression dataset and experimental results demonstrate the superiority of the proposed approach to NMF with different initialization schemes, in terms of relative error, robustness, sparsity and orthogonality.
关键词：Nonnegative Matrix Factorization; Conditional Random Fields; Initialization
