Research On Face Recognition Technique Based On Canonical Correlation Analysis

Face recognition has had its resurgence due to the development of new techniques like face-tagging. In particular, challenges of real world face recognition have been proposed. Feature extraction is one of the most basic questions in Patten Recognition research,for face recognition tasks, extracting the effective facial features is a crucial step. Recently, with the development of feature fusion technology, the Canonical Correlation Analysis method, which will integrates the two feature vectors into a eigenvector with more identifying information, is developing rapidly. In this paper, some algorithms of combined feature extraction is improved by mixed Canonical Correlation Analysis, CICCA algorithm is proposed for color face recognition.Traditional face recognition techniques have mainly focused on the gray level image rather than the color images. However, in real world, most pictures captured by cameras are color ones, whose components contain critical discriminative information for pattern recognition. For color face recognition, we separately perform principal component analysis (PCA), linear discriminant analysis (LDA) and maximum scatter difference (MSD) discriminant analysis on the red (R), green (G) and blue (B) component of color face images to extract three sets of discriminant features. To make full use of the complementary information among R, G and B components, CICCA maximizes the correlation among the three sets of discriminant features, and extract their canonical correlation features for classification. Experiments were done on the AR and FRGC-2 color face database. The results validate the effectiveness of the proposed approaches, which outperform relative methods in classification performance.LDA, Direct-LDA, DCV belong to supervised linear discriminant analysis, and have the same Physical meaning and the similarity process of feature extraction, but as they use the different discrminant criterion, the eigenvectors reflect the different information about the areas of face images. In this paper, LDA, Direct-LDA, DCV are first adopted to extract three sets of features in the same pattern space, respectively. The three variable canonical correlation analysis method is then used to fuse the three sets of features obtained inorder to to derive more effective canonical discriminant features, so as to the unsupervised method:PCA, LPP, SPP.