The partial least squares (PLS) regression is a novel multivariate data analysis method developed from practical applications in real word. In this paper, we first present two new PLS modeling methods (OPLS and COPLS) according to different constraints, and then discuss the two methods theoretically. Based on the idea of PLS model, a new face recognition approach is proposed. The process can be explained as follows: extract two sets of feature vectors from the same pattern, and establish PLS criterion function between the two sets of feature vectors; extract two sets of PLS component (feature vectors) of the pattern by the proposed algorithm, and constitute correlation double-subspace; finally, a serial classifier on the correlation double-subspace is designed, and used in pattern classification. Experimental results on the Yale face image database show that the face recognition approach in this paper is effective.