Face Recognition Based On Canonical Correlation Analysis

In the domain of biometric recognition, face recognition arouses interest amongthe people because of its non-invasive, substantivity, userfriendliness, which hasbecome a very active research subject in the field of pattern recognition. Up to nowthere are a lot of face features extracted by single criterion in the face recognitiondomain. The features can be extracted by optimized combination within themeffectively, which are advantageous for classification and are beneficial to improvethe identification performance of the system. Theoretical frame of the CanonicalCorrelation Analysis (CCA) provides tremendous support for the feature fusion. Sothis paper mainly concludes research on the face recognition using the feature fusionbased on CCA. The major contributions of the paper are as follows:1. On the basis of2DMSD method, firstly this paper gives a new method calledextended2DMSD (EX2DMSD) by redefining the within class scatter matrix and thebetween class scatter matrix. From the view of numerical computation, we know thatthe features extracted by2DMSD mainly contain the row direction information offace, meanwhile the features extracted by EX2DMSD consist of column directioninformation of face, there are complementarity between these two features. So thispaper fuses them by CCA, which can overcome the limitation of single feature in theface recognition throuth making use of row and column direction information of face.The experimental results on the ORL, Yale and AR face databases show theeffectiveness of the proposed method.2. The research shows that the mirror sample of face image can reflect the variationsof perspective, pose, and illumination. So fusing the features extracted from originalimage and mirror image can improve the recognition performance. For this reason, anew face recognition method based on the symmetry CCA feature fusion is proposedby introducing the mirror sample into the CCA. The experimental results illustratesthat the proposed method improves the recognition performance to some extent.3. In order to overcome the problem of LDCCA whose covariance matrice is lack ofclass information, a method called GLDCCA is proposed by introducing the classinformation into the covariance matrix of LDCCA, which can not only preserve thecharacters of LDCCA, but also can keep the discriminate features of the samples.The experimental results on the ORL and Yale face databases show that GLDCCAhas higher recognition rate than the CCA, GCCA, DCCA and LDCCA methods.