Research On Image Classification Based On Nonlinear Kernel Selfadaptive Learning

Kernel learning is an important research topic in the machine learning area,and some theory and applications fruits are achieved and widely applied in signalprocessing, and data mining areas. The nonlinear problems are solved at large withkernel function and system performances such as recognition accuracy, predictionaccuracy are largely increased. However, kernel learning method still endures a keyproblem, i.e., kernel function and its parameter selection. This report aims to supplythe research idea of solving kernel selection problem through the search on kernelself-adaptive learning and its application on3D fragment patching and imageclassification and other areas. The main contents are shown as follows.On the training samples number and kernel function and its parameter enduredby Kernel Principal Component Analysis, this report presents one-class supportvector based Sparse Kernel Principal Component Analysis(SKPCA). Moreover,data-dependent kernel is introduced and extended to propose SKPCA algorithm.Firstly, the few meaningful samples are found with solving the constraintoptimization equation, and these training samples are used to compute the kernelmatrix which decreases the computing time and saving space. Secondly, kerneloptimization is applied to self-adaptive adjust the data distribution of the inputsamples and the algorithm performance is improved based on the limit trainingsamples.On the nonlinear feature extraction problem endured by Locality PreservingProjection (LPP) based manifold learning, this report presents kernel self-adaptivemanifold learning. Firstly, the traditional unsupervised LPP algorithm is extended tosupervised learning and kernelized. Then, kernel self-adaptive optimization is usedto solve kernel function and its parameters selection problems of supervisedmanifold learning, which improves the algorithm performance on feature extractionand classification.On parameter selection problem endured by traditional kernel discriminanntanalysis, this report presents Nonparameter Kernel Discriminant Analysis (NKDA)and this method solves the performance of classifier owing to unfitted parameterselection. On kernel function and its parameter selection, kernel structureself-adaptive discriminant analysis algorithms are proposed and testified with thesimulations.Kernel self-adaptive learning algorithm is applied to the practical system offace recognition systems. Combining Gabor wavelet analysis, this report present akernel adaptive learning machine based face recognition and implement simulation to testify the algorithm on the public databases.Researches show that kernel function and its parameters have the directinfluence on the data distribution in the nonlinear feature space, and theinappropriate selection will influence the performance of kernel learning. Researchon self-adaptive learning of kernel function and its parameter has an importanttheoretical value for solving the kernel selection problem widely endured by kernellearning machine, and has the same important practical meaning for the improvingof kernel learning systems.