Research On Discriminative Canonical Correlation Analysis Based On Semi-supervised Multi-view Data

In recent years,pattern recognition as an important part of artificial intelligence,is widely used in the fields of image recognition,speech recognition,and character recognition.Canonical correlation analysis is a classic algorithm for pattern recognition,which is mainly about the correlation between two views.However,CCA can not take advantage of the class information of the sample,which limits the actual performance of dimensionality reduction.Discriminative canonical correlation analysis is developed based on CCA,and its extracted features can represent the discriminant structure of the sample.However,DCCA can not make use of numerous unlabeled sample information,so it may perform poorly in practical applications.Semi-supervised learning can not only use a small number of labeled samples,but also use a large number of unlabeled samples to improve the recognition performance.Therefore,this paper is based on DCCA in the semi-supervised data scene.The main work can be summarized as follows:(1)This paper wants to use a large number of the unlabeled information to describe the manifold structure of high-dimensional data so as to improve the recognition performance.Therefore,the paper proposes a algorithm which called Manifold Regularization Discriminative Canonical Correlation Analysis(MRDCCA)for semi-supervised data.MRDCCA can use labeled samples to preserve the discriminant information,and estimate the intrinsic geometric structure of data with both labeled and unlabeled samples by introducing the Laplacian regularization terms.Therefore,the algorithm can not only use the advantages of the DCCA,but also preserve the manifold structure of the data.Experimental results on Multiple Features database and face databases(Yale and ORL)show the proposed approach provides a better recognition rate.(2)By introducing the idea of Principal Component Analysis(PCA),and combining PCA with DCCA algorithm with two balance parameters,this paper proposes a algorithm which called Semi-supervised Discriminative Canonical Correlation Analysis based on Principal Component Analysis(PSDCCA).PSDCCA can smoothly bridge the eigenvalue problem of PCA and DCCA.On the one hand,we can make full use of the discriminant information of samples through the DCCA model.On the other hand,we can also utilize PCA algorithm to detect the distribution of spatial data,improve the recognition performance by using a large number of unlabeled samples.The experimental results on several data sets prove that PSDCCA can improve the classification and recognition performance by using a small amount of labeled information and a large number of unlabeled information.(3)In order to make full use of numerous unlabeled sample information,this paper applies Discriminative Canonical Correlation Analysis to semi-supervised data in two stages(DCCA_two).This method can first take advantage of the DCCA to consider the labeled information of the sample,thus maximizing the correlation between samples in the same class.Then it can use LPP algorithm to process a small number of labeled sample information and a large number of unlabeled sample information,so that the point of the nearest neighbor in the original space is still near after reducing the dimension,which better retains the manifold structure hidden in the data.Some experimental results prove that this method has a better recognition effect.