Multi-view Analysis Dictionary Learning For Pattern Classification

Pattern classification is very important for a wide range of applications such as intelligent monitoring and artificial intelligence.In many applications,people usually notice that a certain scene often uses different views to analyze the same object and produce different features.Although many existing multi-view learning strategies have successfully utilized the correlation among views,the joint representation of multi-view learning is still an insufficient problem.At the same time,there is a widespread problem of label missing in data,and the quantity of labeled data is much smaller than that of unlabeled data.In some classification tasks,dictionary learning has shown its excellent performance.As a new research direction,analysis dictionary learning has its theoretical significance and application potential.However,analysis dictionary learning method is lack of research under the new trends such as multi-view problem,semi-supervised training and so on.In this paper,two analysis dictionary learning methods are proposed,and the effectiveness of the proposed methods are verified by experiments.The main contributions of this thesis lie in two parts:(1)A multi-view analysis dictionary learning method MVADL with fully supervised setting is proposed.In this method,a well-state regular constraint and marginalized label strategy is introduced to make the model more discriminating and robust,so as to effectively deal with the correlation among views.Comparative experiments on the standard database verify that the effectiveness of MVADL.(2)A semi-supervised multi-view analysis dictionary learning method is proposed,namely SMvADL.This method makes use of the class consistency information of the data under noises.A mean teacher strategy is utilized to realize the semi-supervised function.In addition,in the case of batch training,the norm regularization of dictionaries and sparse constraints on coefficients enable the model to obtain good parametric solutions.Experiments on several standard databases prove the validity of the method in pattern classification.