Research On Multi-View Classification Algorithms Based On Dictionary Pair Learning

With the increasing diversification of data representation,more and more researchers are paying attention to how to effectively mine and make use of information from multiple perspectives of data for more accurate classification.Dictionary learning as a kind of efficient representation learning technology is widely used in single-view and multi-view recognition,but nowadays most single-view dictionary learning and multi-view dictionary learning algorithms are using one type of dictionary(synthesis dictionary or analysis dictionary),Therefore,based on the dictionary pair learning framework,this paper focuses on single-view and multi-view recognition,and studies classification algorithms with better recognition performance.The research results obtained are as follows:First of all,the traditional dictionary pair learning method not only doesn't combine dictionary learning with classifier learning but also doesn't consider the discriminability and robustness of synthesis dictionary.So we propose a method of discriminative dictionary pair learning based on label embedding and low-rank constraint(LELR-D2PL).By using fisher discriminantive constraint and low-rank constraint on the synthesis dictionary,which not only makes the systhesis dictionary atoms of different categories have a high degree of difference but also makes the whole systhesis dictionary learn good sparse representation from noise data.At the same time,we use the label of synthesis dictionary atom to make constraint which makes coding coefficients of different categories have obvious structural difference.For classifier learning,LELR-D2PL adopts marginalized strategy to adapatively learn regression targets,which ensures the discriminability of the classifier while ensuring the flexibility of the learning.The experimental results on multiple datasets fully demonstrate that LELR-D2PL has better classification performance than traditional dictionary learning methods.Secondly,for most current multi-view dictionary learning algorithms,the diversity information and correlation information of multi-view data cannot be fully utilized.We propose two multi-view dictionary pair learning methods that use different fusion methods,namely,multi-view dictionary pair learning method based on block-diagonal representation(BDR-MVDPL)and multi-view dictionary pair learning based on label embedding and low-rank constraint(LELR-MVDPL).Both BDR-MVDPL and LELR-MVDPL use dictionary pair to obtain discriminant coding coefficients for each view data.Considering the multi-view data sharing label,BDR-MVDPL adopts the feature fusion method to concatenate the coding coefficients of different views,and regress the concatenated coding coefficients to the corresponding label vectors,so that both the diverse information of multi-view data and the correlation of multi-view data are considered.LELR-MVDPL adopts the method of decision fusion to conduct classifier learning within each view,make full use of multi-view data diversity information to make decisions within the corresponding view,and then make a final category prediction by considering the decision opinions of all views.We conducted a classification performance test on the proposed algorithms in multiple public datasets.The experimental results fully show that the proposed algorithms can effectively improve the classification accuracy.Finally,aiming at multi-view human behavior recognition,this paper uses LELR-D2PL as the basic framework and adopts decision-layer fusion to propose a multi-view discriminant dictionary pair learning method(LELR-MVD2PL)based on label embedding and low rank constraints.The corresponding experimental results show that this method can effectively deal with multi-view human behavior recognition.