The Study, Based On Sparse And High-end Classification Of Mode Samples

In order to reduce redundancy of training samples and extract the high dimension feature, we present some arithmetic about samples sparsification and classification based on high order correlation in the paper. We respectively applied these arithmetic to the handwritten numbers and the simulated airplane data(including C130, E2C, F16, IDF, J8II, Mirage2000, SR71, Su27). The results proved the credibility and effect of arithmetic. The specific contents contain:1. Summarized the background, significance and aim of the sparsification and recognition technologies, mainly in the high dimension feature extraction and classification. Moreover, summed up the general situation of the handwritten numbers and HRRP recognition.2. Concluded the characters and pretreatment technologies of the handwritten numbers and radar HRRP(high resolution range profile). This part includes pixel average, distance variety, uniform and so on. Mainly searched three methods: PCA(Principal Component Analysis), KPCA(Kernel-based Principal Component Analysis), and SKLPP(Supervised Kernel Locality Preserving Projections).3. When there were so many similar training samples that can hardly compute, it should restrict the real time implement of the recognition. For this reason, this paper searched some arithmetic about sparsification, mainly in SBayes, SVM(Support Vector Machine), LS-SVM(the Least Square Error Support Vector Machine),KNR(Kernel-based Nonlinear Representor), sparsification based on high order correlation, and then estimated the parameter of classifiers.4. Searched the high order correlation applied to recognition. In order to extract the high dimension feature, we propose in this paper to seek for the correlation coefficient instead of the high order statistics and apply it to sparsification of training sample set and nonlinear classification with radial basis kernels.5. Finally,through the experiments based on Matlab, the results on both handwritten numbers and simulated airplane data show comparatively good performance of the method. We summarized the conclusion and prospected some aspects which are worth to be deep in.