Sub-pattern Canonical Correlation Analysis With Application In Face Recognition

As a classic Multivariate Data Analysis method, Canonical correlation analysis (CCA) extracts features through studying the relationship between two groups of variables. Recently, It has begun to be applied in pattern recognition, but there are some typical disadvantages in small sample size (SSS) problem such as face recognition: (1) CCA fails to be directly applied, due to the singularity of the covariance matrices of its two groups of features caused by the SSS problem; (2) The nonlinear problem can not be well described for its globally linear property in nature; (3) It is short of the robustness to local variants.Nowadays, with the rising of sub-pattern methods, multimodal recognition and other new pattern recognition methods, CCA gets new idea and means to solve these problems. Enlightened by our previous sub-pattern PCA (SpPCA), we present in this paper the sub-pattern canonical correlation analysis (SpCCA). By maximizing the correlation between the local and global features of original samples, this method can not only fuse local and global features well but also eliminate the redundant information among the features.Combined with the sub-pattern method, SpCCA avoids the SSS problem, better formulates the nonlinear face recognition problem and enhances the robustness to the local variants through voting. At present, we can get unimodal data in face recognition usually, when using the multimodal CCA, we generally found multimodal data from unimodal data, then CCA. However, CCA is similar to Linear Discriminant Analysis (LDA) in essence while using hard-label based CCA on multimodal data. And LDA as a unimodal method is born for classification, it has excellent classification performance. So, through replacing CCA by LDA in SpCCA, we present Sub-pattern Linear Discriminant Analysis (SpLDA) to simulate SpCCA. Experiments on AR and Yale shows the performance of SpCCA and SpLDA is comparative, and compared with PCA+LDA, DCV, PCA+CCA, SpPCA and Aw-SpPCA, our methods are more stable, robust and effective.Further, we present overlapped sub-pattern partitioning method to improve Aw-SpPCA, SpCCA and SpLDA. The experiments show it works on all three methods in different degree.