The success of support vector machine depends on the kernel function, which directly affects the performance of SVM. Therefore, to improve the generalization of SVM, we will study the selection of kernel function. The multi-scale kernel method is one particular type of multiple kernel method which combines multi-scale kernels through a multi-kernel learning framework. It has the capability of generalizing not only the scattered region of a training set very well but also generalizing the dense region of data sets very well. Inspired by the advantages of the multi-scale kernel learning method, we applied kernel centered polarization to construct an optimization problem which was used to learn the multi scale kernel function and select the optimal parameters. A thorough analysis and proofs are provided. Experimental results show that the proposed kernel learning method and algorithm are reasonable and effective and have very good generalization performance.