Research On Multi-view Latent Correlational Feature Extraction And Its Application

With the rapid development of data acquisition technology,the scale of data increases rapidly.In the fields of pattern recognition,machine learning,and computer vision,the same set of objects can be described by different features,and this kind of data is called multi-view data.Since each view generally has a higher dimension,the direct use of raw multi-view data to learn related tasks often leads to poor learning performance and other problems.Therefore,it is necessary to carry out joint dimensionality reduction for high-dimensional multi-view data.Partial Least Squares(PLS)and Canonical Correlation Analysis(CCA)are two classical multivariate statistical analysis algorithms,which are very suitable for multi-view feature extraction.Therefore,according to the different application scenarios,this paper takes multi-view latent correlational feature extraction as the starting point,combines PLS and CCA algorithms with supervised learning,the idea of fractional order and deep network,and establishes a set of algorithm system for image recognition.The main research work and results of this paper are as follows:(1)Double Directional Two-Dimensional PLS(D2PLS)algorithm is proposed.This algorithm takes into account both the row information and the column information of the matrix image,makes up for the deficiency of the existing Two-dimensional PLS(2DPLS)algorithm which does not take into account the row information of the matrix image,and can extract features from the data of two views.Experiments on AR,Yale,and AT&T face databases show that D2PLS has a significant improvement in recognition performance compared to traditional 2DPLS.(2)Fractional-order Labeled Multiple CCA(FLMCCA)and Fractional-order Discriminant Multiple CCA(FDMCCA)algorithms are proposed.FLMCCA and FDMCCA add label information on the basis of CCA,and reduce the deviation of covariance matrix caused by noise disturbance or limited training samples by introducing fractional order parameters.In addition,both algorithms can be applied to more than three views.Experimental results on AR,CMU-PIE,AT&T face database and UCI handwritten digital image database show that the features extracted by FLMCCA and FDMCCA are more discriminative and effective than those extracted by traditional algorithms.(3)Deep Supervised Multiple CCA(DSMVCCA)algorithm is proposed.Based on Deep CCA(Deep CCA),this algorithm extends it to multiple sets and adds category label information.The features extracted by DSMVCCA algorithm can not only reveal the nonlinear dependence among multi-view data,but also have strong discriminative power.Experimental results on MNIST handwritten digital database and AR face database show that the DSMVCCA algorithm has a good recognition effect.