The Research Of Multi-view Latent Association Mining Based On Image Set And Non-negative Matrix Factorization

With the rapid development of computer technology,the capability of data collection and storage has been improved greatly.In many real-world applications,often confront with data representation in multiple views,multi-view learning has attracted widespread attention and become a hot research topic in the machine learning and pattern recognition.Multi-view learning focuses on the mining and integrate the latent consistent and complementary information.There are two research problems in the research of algorithms based on multi-view latent association mining,namely multi-view supervised learning and multi-view unsupervised learning.In multi-view supervised learning,human action recognition from videos is a challenging research topic,which has received a significant amount of attention from the research community due to its wide range of application such as intelligent surveillance system and human machine interface.In multi-view unsupervised learning,multi-view clustering is a fundamental and hot research topic,it has mainly used in many fields such as text clustering and image analysis.Therefore,this paper will study multi-view learning from the above two perspectives.The research work in this paper mainly includes the following two parts: 1)The research of multi-view human action recognition based on supervised multi-view latent association mining.2)The research of multi-view clustering based on unsupervised multi-view latent association mining.1)A multi-view human action recognition algorithm based on adaptive fusion and category-level dictionary learning model(AFCDL)is proposed.In this algorithm,the adaptive weight is learned for each view for mining the latent complementary relationships among different views.In order to guide the recognition process of the test set and improve the recognition accuracy,the induced set for each category is built,and the corresponding induced regularization term is designed for the objective function to guide the dictionary learning and the reconstruction of query set.Meanwhile,the consistency regularization term is used to ensure that the sparse representation coefficients of the training subset and the induced subset tend to be as consistent as possible.Two classification schemes are designed to predict the label of the test samples,minimum error scheme and linear SVM scheme.Furthermore,alternating minimization method is employed to optimize the objective function.The experimental results in UCLA,IXMAS,CVS-MV-RGBD-Single and CVS-MV-RGBD-Double datasets show that AFCDL has significant advantages over the existing multi-view human action recognition algorithm.2)A multi-view clustering algorithm via consistent and specific non-negative matrix factorization with graph regularization(MCCS)is proposed.In this algorithm,the consistency among all multi-view data and view-specific information in each view data are simultaneously discussed in the MCCS algorithm proposed in this paper.A novel consistent and specific non-negative matrix factorization expression is proposed.And the manifold regularization is embedded into the objective function to preserve the intrinsic geometrical structure of original data space.Meanwhile,a disagreement term is employed to make these view-specific coefficient matrices to further towards a common consensus and ensure the multiple views have the same underlying cluster structure.Furthermore,multiplicative update algorithm is employed to optimize the objective function.The experimental results in BBC,BBCSport,20 NGs,Wikipedia and Handwritten datasets show that MCCS has significant advantages over the existing multi-view clustering algorithm.