An approach for directly extracting features from matrix data and its application in face recognition

By formulating two-dimensional principle component analysis (2DPCA) as a mathematical form different from the conventional 2DPCA, we present theoretical basis of 2DPCA and show the theoretical similarities and differences between 2DPCA and PCA. We also show that 2DPCA owns its decorrelation property and the feature vectors extracted from matrices are uncorrelated. We use the proposed mathematical form to show that 2DPCA is the best approach for directly extract features from matrices. We also present in detail 2DPCA Schemes 1 and 2, two schemes to implement the proposed mathematical form. The two schemes transform original images into different spaces, respectively, 2DPCA Scheme 1 enhances the transverse characters of images, whereas the second scheme enhances vertical characters of images. We propose a feature fusion approach for achieving better recognition results by combining the features generated from the two schemes of 2DPCA. The proposed fusion approach is tested on face recognition tasks and is found to be more accurate than both 2DPCA Scheme 1 and 2DPCA Scheme 2.