Globalized And Localized Multiple Empirical Kernel Learning And Its Application

Multiple Kernel Learning (MKL) is demonstrated to boost classification performance through using multiple kernels rather than a single fixed kernel. MKL has two kinds of k-ernel mapping forms, including the traditional implicity kernel map and the empirical kernel map, which are called Multiple Implicity Kernel Learning (MIKL) and Multiple Empirical K-ernel Learning (MEKL), respectively. In this paper, we focus on MEKL and develop a novel Globalized and Localized MEKL (GLMEKL), which has the following contributions.Firstly, in this paper, a new method of Canonical Correlation Analysis(CCA) named Glob-alized and Localized CCA (GLCCA) is designed, by means of researching CCA and some existing improved CCA-based methods. GLCCA can get both the global and local structural information between two group samples, and construct the fusion samples reflecting the struc-tural characteristics of the original ones. The fusion samples are in favour of building a better classifier. Then, the GLMEKL framework is proposed by combining GLCCA and MEKL. It is known that the existing MEKL or MIKL deals with the data mapped from different kernels in the classifier level. By contrast, GLMEKL mainly works in the feature level by obtaining both the global and local structural information between different empirical kernel mapped s-paces at the same time. It will improve the classifier performance through the fusion samples in classification task.Secondly, We used three types of data sets(UCI datasets,face images datasets and multi-view datasets) in GLMEKL experiment by comparing with the existing related methods. A large amount of charts and text listed as well as the factors affecting the GLMEKL perfor-mance. The experimental results validate that the GLMEKL can not only inherit the advan-tages of MEKL, but also adopt the structural information from different mapped spaces, so it can improve the performance of classifier.Finally, a face recognition system based on MATLAB platform is designed to show GLMEK-L realization process clearly. The system framework is based to demand analysis and the prin-ciple of software engineering, also combined with input and output environment and process control for the face recognition system. After that, we completed the coding work. The system functions are introduced in detail, as well as the interface show by some screenshots.