(Dpls)-P-2: A Novel Bilinear Method For Facial Feature Fusion

Two-dimensional partial least squares (2DPLS) is an effective two-view data analysis technique. However, conventional 2DPLS only takes into account the column information of two-dimensional images. In this paper, we simultaneously consider the column-wise and row-wise information of two-dimensional face images. We first propose a row-based two-dimensional PLS (r(2)DPLS) approach and then further present a novel double-directional PLS (D-2 PLS) method. The proposed D-2 PLS method can be optimized by two eigenvalue subproblems. Experimental results on the AR, Yale, and AT&T face databases show that our (DPLS)-P-2 method can overall achieve better recognition accuracy than existing related methods.