Research On Multi-view Feature Fusion Methods

With the development of the Internet technology and its applications in daily life widely and deeply,a large amount of data have been produced constantly.Thus the appearance and convenience of all kinds of media are suffered from new challenges for requirements proposed by people.Multi-modal and multi-view data show the different characteristics of an object from multiple data sources or multiple perspectives which have attracted more and more attentions under the background of big data.For such type of data,if we could find the appropriate methods for feature fusion,we should improve the efficiency and effectiveness of relative tasks in a great degree.In fact,multi-view feature fusion also brings new ideas for dealing with small-scale sample problems such as extracting multi-type features from the small-scale samples which is equivalent to the expansion of samples essentially.In this dissertation,we focus on multi-view feature fusion methods and list the research works as follows.(1)A multi-view feature inverse fusion method based on support value transform(SVT)is proposed.The proposed method uses SVT to extract the multi-view fea-tures of the face image and performs inverse fusion to obtain the low frequency representation(LFR)of the face image.The proposed method directly works on the two-dimensional image matrix,which avoids the information loss when trans-forming the image matrix into one-dimensional vectors before the image feature extraction.Experimental results on two publicly available face databases ORL and UMIST show that the proposed method can provide better recognition results for face images or small-scale images under different illumination conditions.(2)A multi-view Lie group feature fusion based on region covariance method is present in which the region covariance is taken as a kind of features for images,that is to say the covariance of several statistical features of the concerned image region is used as the descriptor of the region.The computational complexity of region covariance is small and the computational cost of region covariance is independent of the image region size.In addition,region covariance can avoid the influences of image rotation and illumination variations.However,covariance matrix is not an element of Euclidean space.Thus,we cannot use the existing classical machine learning algorithms to deal with it.To remedy it,this dissertation proposes a multi-view Lie group feature fusion method based on region covariance.This method fully uses of the ability of complex data representation and distance calculation of Lie group,and can effectively solve the problem of feature fusion and classification of complex and high dimensional image data.Experiments on a variety of public datasets show that the proposed method has good results.(3)A robust multi-view feature fusion based on collective non-negative matrix fac-torization(CNMF)method is proposed.In many practical applications,the data usually contain noise,which would affect the performance of multi-view learn-ing.Be inspired by the successful application of collective matrix factorization(CMF)in multi-view learning,we propose a robust collective non-negative matrix factorization(RCNMF)model which could fuse the features and denoise features simultaneously.Thus,the proposed method can find the fusion subspace which is not affected by noise as far as possible.The proposed method is verified by experiments on several public datasets.(4)A supervised orthogonal discriminant projection based on double adjacency graphs is proposed.The classical multi-view subspace algorithms assume that each view arises from the same hidden subspace.In order not to depend on this assumption,this dissertation proposes a supervised orthogonal discriminant projection based on double adjacency graphs(SODP-DAG)to reduce the dimension of data.Accord-ing to the Fisher criterion,the proposed method takes into account the importance of between-class and within-class scatters for images classification and solves the problem in the orthogonal discriminant projection method which only considers minimizing the local scatter.By using double-adjacency graph,the local scatter can be further divided into the local between-class scatter and within-class scatter,the proposed method could find the valid discriminant projection directions in the case of preserving the local structure of the original data.In addition,four schemes are designed using the heat kernel function to construct the weight matrix so as to further deal with the fusion of multi-view features.To verify the effectiveness of the proposed approach,we compared it with other related methods on a variety of pub-licly available datasets.Experimental results show the feasibility and effectiveness of our method.