Research On The Classification Method Of Heterogeneous Data Based On Feature Fusion

Pattern recognition is a method of automatically dividing samples according to the characteristics of sample data by using mathematical modeling.Pattern recognition has been widely used in computer-aided diagnosis,image segmentation,object detection and many other fields.In the traditional pattern recognition methods,the features of a single mode are generally used to recognize samples through some machine learning algorithms.However,the features of a single mode cannot contain all the characteristics of the samples..At the same time the data of the first modal often have uncertainty or data loss due to the environmental impact or data transmission.The uncertainty and lack of data will not only bring difficulties to pattern recognition,but also reduce the accuracy of the classification model.Multi-source heterogeneous data is obtained through different channels,and the data obtained through different channels contains different characteristics of samples.The accuracy and robustness of the classification model can be improved by the complementary information of multi-source heterogeneous data.How to effectively use the relationship between different modal features to establish an effective classification model is one of the main problems faced by researchers.Feature fusion of multi-source heterogeneous data can effectively use the specific information in multi-source heterogeneous data,which is an effective method to balance the relationship between multi-source heterogeneous data.There are still many difficulties in the current feature fusion methods:(1)the uncertainty in the process of data acquisition leads to poor data quality;(2)Data are heterogeneous;(3)How to balance the relationship between heterogeneous data and within data to build an effective classification model.To solve the above problems,we have done the following work in this paper:1)Aiming at the problem of data missing in multi-modal heterogeneous data,a variety of methods are used to complete the missing data.In order to avoid the influence of noise caused by data completion,the mathematical modeling of the missing data and complete data in Cartesian coordinate system is proposed by using tensor outer product.This method does not need to complete the data.2)In order to better represent the relationship between multi-source heterogeneous data,feature weighting and discriminant correlation analysis are used to model the relationship between multi-source heterogeneous data.At the same time,in order to avoid the loss of information between data,the method of low rank representation is improved.High-order Laplacian matrix is introduced to learn the internal relationship between multi-source heterogeneous data based on the low-rank representation.The main innovations of this paper are as following:(1)The relationship between multi-source heterogeneous data is modeled to get the fusion matrix,which is used for classification to avoid the influence of noise and redundant information brought by the original data.(2)In order to avoid the influence of noise caused by data completion,tensor outer product is used to model the complete unimodal data and missing unimodal data in Cartesian coordinate system.(3)High-order Laplacian matrix is introduced to learn the high-order relations between multi-source heterogeneous data based on low-rank representation,and more information is obtained through the learning of high-order relations.