Research On Canonical Correlation Analysis Based On Label Sensitivity And Structure Constraint

With the progress of social economy,science and technology,image analysis and recognition are more and more widely used in life.The actual collected image data is usually complex high-dimensional data.Extracting distinguishing features from high-dimensional data has become a hot research topic.At present,there are still some problems and limitations in the research of feature extraction,such as ignoring the class label supervision information in the sample,affecting the discrimination ability,the small sample and noise of the sample lead to excessive redundancy of features,and the multi-source layer characteristics of the target cannot be found by dual view feature extraction.Aiming at these problems,this thesis focuses on the embedding of discrimination information,orthogonal redundancy reduction and the expansion of multiple sets,combined with canonical correlation analysis theory,a series of cross view feature extraction methods are proposed.The main contents and contributions of this thesis could be summarized as follows:(1)A feature extraction method based on Discriminant sensitivity is proposed.The class identification information of samples is the effective supervision information in feature extraction,which can improve the feature discrimination.The popular supervision feature extraction methods can not retain the structural information of class features.To solve this problem,based on the class label identification of sensitive information,this thesis deeply studies the identification structure under the cross view canonical correlation analysis framework,proposes a feature extraction method based on identification sensitivity,explores and retains the global structure of supervision information,enhances the class separation of relevant features with the help of inter-class and intra-class dispersion structure,and verifies the validity of the proposed methodology in experiments.(2)A canonical correlation analysis method for discriminating redundancy reduction is proposed.In practical application scenarios,only considering class information and ignoring redundant information will make the extracted features have great defects.To overcome the above deficiency,based on the orthogonal structure constraint theory and combined with the identification sensitive information,this thesis proposes a typical correlation analysis method of identification redundancy reduction,which can better preserve the reconstruction relationship between data,reduce the redundant information of projection matrix,preserve the useful structure of projection matrix and improve the identification ability of fusion features.A large number of comparative experimental results verify the validity of the proposed methodology,further improve the model discrimination and enhance the robustness of the model.(3)An orthogonal canonical correlation analysis method with multiple sets discriminant sensitivity is proposed.In different data sets and different application scenarios,if a group of targets are represented by multiple feature data sets,it can more comprehensively show the multi-source characteristics of this group of targets.Moreover,in the pattern recognition task,the noise problem of small samples and samples has always led to the error of sample covariance matrix.To solve the above problems,combined with orthogonal constraint criterion and supervised dispersion structure,this thesis proposes a multi set discriminant sensitive orthogonal canonical correlation analysis method,which reduces the redundant information of feature extraction and realizes the discriminant extraction of more than two groups of features.In this thesis,the experimental comparison on multiple data sample sets is completed and the experimental results are evaluated and analyzed.The results show that this method has strong recognition ability in image recognition.Figure [7] Table [12] Reference [59]...