Multiple Kernel Support Vector Machine And Its Application In Face Recognition

Support vector machine (Support Vector Machine, SVM) is a machine learningmethod which bases on kernel function. It was widely used in pattern recognition、machine learning and other fields. The kernel function or parameters of kernel functionhas the significant influence to the generalization ability of the support vector machine.But in some complicated cases, researchers find that the kernel machines with a singlekernel function can not meet some practical requirements such as heterogeneousinformation or unnormalised data, large scale problems, non-flat distribution of samples,etc. In recent years, some researcher propose multiple kernel support vector machine,they consider the combination of kernel functions for better results. Assigning the sameweights to a kernel in different regions of the input space can not reflect the localinformation. Localized multiple kernel learning(LMKL) is composed of a kernel-basedlearning algorithm and a parametric gating model to assign local weights to kernelfunctions.It is an attractive strategy for combining multiple heterogeneous features interms of their discriminative power for each individual sample.But the gating functionof LMKL algorithm the is redundant, hence, this paper proposes an improved localizedmultiple kernel learning algorithm.The technique of Tikhonov regularization was usedin this paper,moreover, this paper proposes an new gating function with arbitraryl pnormconstraint.The feature extraction of face image is a premise of recognition work, Featureextraction directly affect the results of the recognition, and the design of classifier is thecore work of face recognition, so, this paper focuses on the research feature extractionand classifier design of face recognition. In the feature extraction stage, this papermainly used principal component analysis and the wavelet transform method. In theclassifier design stage, using the simple simpleMKL、LMKL and ILMKL as theclassifier respectively. Evaluations of ILMKL on both benchmark machine learningdata sets (UCI data sets) and ORL data sets have shown to achieve state-of-the-artperformances.