Research On Feature Extraction Algorithm Based On Deep Canonical Correlation Analysis

In pattern recognition,an object usually has different feature representations,and these different representations can reflect different information or views of the same object.Deep Canonical Correlation Analysis(CCA)is a widely used feature extraction algorithm,whose goal is to learn the complex nonlinear transformation between two data views,so that the obtained feature representations have a high degree of nonlinear correlation,which can be seen as a nonlinear extension of linear canonical correlation analysis,the deep canonical correlation analysis is essentially an unsupervised subspace feature learning algorithm,and the class label information of each sample is not fully considered.In addition,the deep canonical correlation analysis does not have good robustness when faced with small sample problems.To address above mentioned problems,this paper is based on the deep canonical correlation analysis theory.consider the fractional order theory to it,and makes full use of the data label information to guide the learning process,makes deeply researches and constructs the deep canonical correlation analysis theory of supervision information,and it is applied to recognition tasks such as face recognition and speech recognition.The main innovations and research results of this paper are as follows:(1)A supervised deep correlation analysis algorithm(SDCCA)is proposed.The algorithm makes full use of the class label information of the sample,and guides the learning of the parameters,so that the learned feature representation can maintain the correlation as much as possible,and can also ensure the maximum distance between the new projections.The experimental results on the handwritten data set MNIST and the data set COIL-100 show that the features extracted by this algorithm can improve the classification ability of Deep CCA.(2)A complete deep canonical correlation analysis(Complete deep CCA)is proposed.The algorithm redefines a new correlation calculation method.On the basis of the original method.this method not only considers the correlation between the same samples,but also considers the correlation between different samples of different samples in the same category.This makes the calculation of correlation more comprehensive.Compared with the original deep canonical correlation analysis algorithm in the handwritten data set FFER20I3 and speech data set RAVDESS,the recognition rate is mproved,which indicates the effectiveness of the new algorithm.(3)A fractional order embedding deep canonical correlation analysis is proposed(FEDCCA).In face of small sample problems,there is usually a large deviation between the sample covariance matrix and the real covariance matrix.The algorithm uses the idea of fractional order to re-estimate the covariance matrix of the sample to make it closer to the true value.We call the new estimated covariance matrix a fractional order scatter matrix.The experimental results show that the FEDCCA algorithm further improves the recognition rate of the deep canonical correlation analysis algorithm when the number of samples is insufficient,which indicates that it is meaningful to re-estimate the covariance matrix.