Based on the theory of kernel function and support vector machine, adaptive selective kernel method is investigated in this thesis. An improved algorithm of on line kernel function is proposed by using Naive Regularized Risk Minimization. As time window slide, the new observations can be trained through the algorithm with pruning error minimization and Lagrange factor selection. And so, the adaptive realization of sample choice is resolved and the higher convergence precision can be obtained comparing the present method.Focused on the features and requirements of independent component analysis, the improved algorithm is introduced for human face recognition. And the relative algorithm based ICA-SVM is also developed.The characteristics of algorithms presented here are analyzed by separately using computer simulation and the database existed of human face. The results show us that the algorithms are reliable and feasible.The related software is developed on Matlab6.5 and Labview7.1. |