Research On Feature Selection Method In Multi-Source Target Recognition

Target recognition has been developed from a past research for a single source to a multi-source.Each type of data source is described from different perspectives for the target,which leads to the proliferation of massive data.While expanding the connotation of classification features,it also brings a lot of information redundancy and data correlation,leading to the "dimension disaster",which affects seriously the efficiency of machine learning and the performance of classification algorithm.How to effectively screen and utilize the information provided by various sensors has been an important research value and practical significance.On the one hand,the increase of feature dimension can provide more sufficient information and bring more comprehensive content for subsequent processing.On the other hand,the redundant and irrelevant features in the feature increase accordingly,which not only reduces the learning speed and classification accuracy of the learning algorithm,but also causes the problem of "dimension disaster".Feature selection algorithm is one of the typical methods of massive data high-value information screening.As an effective dimension reduction method,it is widely used in multiple fields of pattern classification and target recognition.Therefore,based on the basic theory of information theory and taking information entropy as the basic measurement strategy,this paper studies the multi-feature selection method for target recognition applications.Firstly,an rMIFS(rankingMIFS)algorithm is proposed to measure the magnitude imbalance of the evaluation criterion function in the traditional feature selection algorithm based on mutual information.The algorithm first sorts the numerical values of relevancy,and then use the serial number to replace the number,which improves the evaluation of the criterion function in the later stages of the feature selection.The validity of the proposed method in the open data set,and the results show that the rMIFS algorithm has improved the integrity of the criteria function in the whole process,which has improved the ability to classify the best children.Then,for the accuracy of the estimation of Shannon mutual information heavily dependent on the probability density function in calculation,a new measurement method-Survival cauchy-schwartz mutual information is introduced,and a scs-rmifs(Survival cauchy-schwartz MIFS)algorithm is proposed.Survival Cauchy-Schwartz mutual information replaces the probability density function with survival function,which can effectively avoid the influence of sample size,artificial parameter setting and other problems on the correlation and redundancy measurement of data set.The effectiveness of the proposed method is verified on the open data set.The results show that the proposed scs-rmifs algorithm improves the accuracy of feature measurement,the most favorable features for classification in each selection is selected,and the classification ability of the selected optimal subsets is improved.Finally,based on the above research,in view of the interested target of the multi-source image fusion detection problem,this paper puts forward a kind of image feature selection and super pixel segmentation algorithm(SLIC)combined classification strategy,added the dimensions of the SLIC distance vector algorithm,and give segmentation tags semantic constraints,pedestrian detection in visible and infrared image fusion for application demonstration,realizes the rapid classification of the target of interest extraction.The effectiveness of the proposed method is verified on the real data set.The results show that the classification effect of the algorithm is improved in overcoming the incomplete segmentation and fragmentation of the target.