Feature Extraction And Fusion Based On Multi-view Correlation Projection Analysis

Due to different data collection technology and processing,the same objects usually have multiple representations from different views.For the same objects,these multiple representations reflect different characteristics,extracting features from them can not only obtain effectively discriminative information,but also eliminate the redundant information to a certain extent in each feature representation.However,the linear and linear projection analysis methods are processing mainly on one feature of the patterns.It is unsuitable to apply it to directly extract and fuse the features of multi-view data.Multi-view correlation projection analysis,including canonical correlation analysis(CCA),multiset canonical correlation analysis(MCCA)and multiset integrated canonical correlation analysis(MICCA),has been widely employed in multiple feature fusion and extraction,and it has obtained good recognition results in the application of pattern classification.In this dissertation,we focus on researching the multi-view correlation projection analysis,in order to learn a consensus pattern based multi-view,which is more accurate than the patterns based on a single view representation.Our work mainly includes the following parts:(1)Sparse regularized semi-supervised discriminative canonical correlation(SrSDCC)is proposed.In practice,when there is no sufficient training samples,the covariance matrix of each class may not be accurately estimated.In oder to solve this problem,we propose a novel method,called semi-supervised method,which makes use of both labeled and unlabeled samples.The labeled data points are used to maximize the within-class correlation and minimizes the between-class correlation,the unlabeled data points are used to estimate the intrinsic geometric structure of the data.Extensive experiments on face recognition demonstrate that the proposed method can effectively enhance the recognition performance.(2)Generalized fuzzy canonical correlation analysis(GFCCA)is proposed to reflect the samples distribution information.Through employing fuzzy k-nearest neighbor method to calculating the membership-degree and takes the membership-degree to characterize the similarity weight between the sample and the center of a class,it effectively utilizes the class label information and improves the effectiveness and robustness of the algorithm.Moreover,for nonlinear separated problems,we extend the kernel extension of GFCCA with positive definite kernels and indefinite kernels.The experimental results on AT&T,Yale and and UCI multiple feature databases show that the GFCCA has more discriminating power and can provide encouraging recognition results in contrast to the state-of-the-art algorithms.(3)Multiset canonical correlation analysis(MCCA)is a powerful algorithm for analyzing linear relationships among several(more than two)sets of features.However,MCCA contains an multivariate eigenvalue problem(MEP),and it is difficult to gain the accurate solution vector,resulting in the constraint of the recognition performance.An inaccurate solution vector sometimes induces instability of MCCA algorithm.In order to solve the problem exists in MCCA,we incorporate power-symmetric successive overrelaxation(P-SSOR)from algorithm angel,named as PssorMCC.Extensive experimental results on the handwritten Arabic numeral,object and face datasets show that the PssorMCC algorithm has good recognition performance.(4)Margin linear discriminant multiset canonical correlation analysis(MLDMCC)is proposed for feature extraction of the high dimensionality multi-view data.Due to MCCA fails to discover the intrinsic geomerical and discriminating structure of multiple data spaces in real world applications.In this paper,we thus propose a novel algorithm based on margin fisher analysis,called MLDMCC,which explicitly considers both discriminative and intrinsic geometrical structure in multiple representation data.MLDMCC combined with the discriminative information of linear discriminant analysis,it can not only express the correlation among multiple feature vectors,but also effectively depict the data between the geometry and discriminative structure.Extensive experiments on both face image databases and object databases demonstrate the effectiveness of the proposed method.(5)For using the class label information to pattern classification and acquiring good results,two supervised MICCA,called discriminative multiset integrated canonical correlation analysis(DMICC)and generalized multiset integrated canonical correlation analysis(DMICC)are proposed.Moreover,in order to fully reveal multiple kinds of characteristics and geometrical structure information of original data in multiple high-dimensional feature spaces,we construct MKMICCA by using the ideas from multiple kernel learning(MKL).MICCA contains the generalized multivariate eigenvalue problem(GMEP),which is more complex than MEP.Based on the multiple eigenvalue problem solving method,this paper proposes a new iterative algorithm to gain the accurate solution vector.The experimental results show that the proposed algorithms not only have good recognition performance,but also obviously outperforms the state-of-the-art algorithms.