Uncorrelated Multi-Set Feature Learning for Color Face Recognition.

Fei Wu,Xiao-Yuan Jing,Xiwei Dong,Qi Ge,Songsong Wu,Qian Liu,Dong Yue,Jing-Yu Yang
DOI: https://doi.org/10.1016/j.patcog.2016.06.010
IF: 8
2016-01-01
Pattern Recognition
Abstract:Most existing color face feature extraction methods need to perform color space transformation, and they reduce correlation of color components on the data level that has no direct connection with classification. Some methods extract features from R, G and B components serially with orthogonal constraints on the feature level, yet the serial extraction manner might make discriminabilities of features derived from three components distinctly different. Multi-set feature learning can jointly learn features from multiple sets of data effectively. In this paper, we propose two novel color face recognition approaches, namely multi-set statistical uncorrelated projection analysis (MSUPA) and multi-set discriminating uncorrelated projection analysis (MDUPA), which extract discriminant features from three color components together and simultaneously reduce the global statistical and global discriminating feature-level correlation between color components in a multi-set manner, respectively. Experiments on multiple public color face databases demonstrate that the proposed approaches outperform several related state-of-the-arts.
What problem does this paper attempt to address?