Research On Cross-view Correlation Feature Extraction Based On Discriminant Sensitivity

With the rapid development of information technology,it brings massive high-dimensional data and redundant information with strong correlation in the original data.It has become a hot research field to extract features with identification ability from high-dimensional data by computer.However,there are still some problems and limitations in the research at this stage.For example,the feature extraction of single view data can not effectively represent the multi-source hierarchical characteristics of a pattern resulting in poor feature recognition ability and weak anti-interference ability,and the feature extraction method based on Gaussian hypothesis can not effectively solve the problems of practical non normal application scenarios,Failed to fully explore the internal structure and category contribution of supervised information,and unsupervised feature extraction methods can not simulate the identification structure and improve the identification ability.In view of these problems,this paper will focus on multi perspective information feature extraction,mainly cross perspective feature extraction methods,using the mutual information of different perspectives,Three multi view feature extraction methods based on Discriminant sensitive canonical correlation analysis are proposed,and comparative experiments are carried out on several different public data sets to verify the universality of the proposed method.The main research contents of this paper are as follows:(1)The discriminant features of single view data feature extraction are limited by the singleness of single view representation,and can not well mine the multi-source characteristics of the target pattern.The single view linear discriminant analysis based on the assumption of normal distribution can not solve the problem of cross view feature extraction and the influence of outliers.In this paper,a cross view feature extraction method through canonical correlation analysis and p-norm enhancement method with truncation function canonical correlation analysis are proposed,which can effectively deal with the problem of feature extraction of multi view data with various distributions.A large number of comparative experimental results verify the effectiveness of p-norm cross view correlation feature extraction method with truncation function,It further improves the discrimination of the model and increases the robustness of the model.(2)Class label information is an effective supervised information,which can effectively improve feature discrimination.Under different data sets and different application scenarios,the popular supervised feature extraction methods fail to consider the contribution rate of different classes,so as to explore the global structure of label information.To solve this problem,this paper proposes a cross view feature extraction method based on label self balance,explores the global structure of supervision information,integrates the adaptive framework,extracts adaptive information from label data,and realizes adaptive weight assignment for category contribution.In addition,an alternative iterative method is proposed to solve the multivariable problem,and a stable optimization solution is obtained.Experiments verify the effectiveness and robustness of the proposed method.(3)Considering that the acquisition of supervision information may not be satisfied in practical application scenarios,the research on unsupervised cross view feature extraction algorithm still has high practical significance.Aiming at the problem that cross view feature extraction algorithm canonical correlation analysis can not train the model through label information with effective clustering function,this paper proposes a cross view feature extraction method of supervised approximation,introduces the farthest neighbor mutual exclusion relationship to simulate the relationship between classes,and achieves the discrimination sensitivity of approximation by minimizing the relationship between classes through geometric preservation.In the unsupervised environment,Weaken the redundant inter class correlation,approach supervised learning and improve feature discrimination.In this paper,a comparative experiment is carried out on public face data set and handwritten data set.The performance evaluation and analysis of the experimental results show that this method has strong recognition ability in image recognition.Figure[12]Table[16]Reference[61]...