Research And Application Of Correlation Filter Based On Data Structure Information

With the development of computer technology,target recognition and target detection have attracted wide attention of researchers in the fields of computer vision,pattern recognition and image processing.Among them,the correlation filter algorithm is one of the current research hotspots.The goal of the correlation filter is to improve the filter's recognition and detection capabilities by training the sample to obtain a filter template that better matches the sample information.Based on the traditional correlation filters,Maximum Margin Correlation Filter(MMCF)algorithm combines Support Vector Machine(SVM)with an equilibrium factor to minimize the mean square error and maximize the minimum Euclidean distance from the hyperplane to the data point.However,the structural information of the sample data is not further used to some extent,which limits the performance of the MMCF.SVM only considers the sample points on the boundary when establishing the classification hyperplane,so that MMCF also inherits this disadvantage.When constructing the filter,only the boundary sample point information is applied,and the overall distribution information of the sample is ignored.It is susceptible to noise samples.At the same time,the within-class structure and the specific distribution of data are not applied to the training,and ignored data structure information when constructing filters.In order to further improve the recognition and detection capabilities of correlation filters,this article will study the above two aspects,the main contents are as follows:(1)Aiming at the problem that the MMCF does not consider the overall sample distribution information when constructing the filter,the within-class divergence in the minimum class variance support vector machine(MCVSVM)is introduced into the correlation filter.The distribution information within the class reduces the differences in the same class,while maximizing the classification interval,and fully considering the distribution of the sample data,and then proposes minimum class variance correlation filter(MCVCF).(2)Aiming at the problem of MMCF ignoring the within-class structure of various types and the specific distribution of data,integrate the correlation filter with the sample weighted adjacency graph,introduce the locality preserving within-class scatter,fully consider the sample distribution information and the within-class flow pattern structure,and maximize the classification within-class and optimize the correlation output,taking into account the sample's category information and structural information,then proposed Minimum Class Locality Preserving Variance Correlation Filter(MCLPVCF).In order to verify the detection and recognition performance of the proposed algorithm,the experimental categories such as object recognition,face recognition,and human eye detection are compared with several classic related filter algorithms under different experimental conditions,and the parameter pairs of the algorithm are discussed.Impact of results.Experiments show that the proposed algorithm achieves better detection success rate and average recognition rate in most cases.