Currentcy Recognition Based On Multiple Kernel Learning Support Vector Machine

Abstract:Support vector machine(SVM) is a new machine learning technique developed on the basis of statistical learning theory. SVM has great advantages on nonlinear, high dimension, and small sample pattern recognition such as paper currency recognition. The choice of kernel function, paper currency feature fusion problem, and reducing the generalization error upper bound of SVM classifier are issues of crucial importance. The main shortcomings of the existing paper currency classifiers based on SVM is that when the SVM is used to solve the paper currency recognition, multiple features are usually integrated but only one kernel function is used. Moreover, loss function ignores the fact that the value of the radius of minimum enclosing ball imposes great influence on generalization ability. This paper undertook great research to solve these problems and puts forward a novel paper currency classifier based on SVM. This new classifier adopts multiple kernel learning and integrates information of radius into loss function to solve three problems mentioned above and improve paper currency recognition.Choosing the most suitable kernel functions for paper currency recognition by SVM is the difficult issue, when multiple features are usually integrated. Kernel functions have important influences on learning ability and generalization ability of SVM. This paper comes up with the idea that using a linear combination of kernels to build multiple kernel learning is an effective solution to the problem of choosing basis kernel functions. Moreover, multiple paper currency features can be fused in the process of training multiple kernel learning SVM through kernel functions. In this way, paper currency feature fusion problem can be effectively solved. Finally, the reduced gradient method, iterative optimization, and sequential minimal optimization(SMO) algorithm are used to find the optimal weight coefficients of basis kernel functions and solve the dual problem of multiple kernel learning SVM. It’s showed in the experiments that multiple kernel learning SVM can make up the limitations of single kernel and there is a guarantee for achieving better generalization ability and learning ability. And the multiple kernel learning SVM is a more efficient solution to multiple paper currency features fusion.Generalization risk of SVM classifier depends on the maximal margin and the radius of minimum enclosing ball, but traditional SVM model ignores the information of radius. This dissertation presents a novel multiple kernel learning model, which integrates information of radius and then optimizes both the maximal margin and the radius to reduce the upper bound of the generalization error of multiple kernel learning classifier and gain better generalization. Primal problem of2-norm soft margin multiple kernel learning SVM including the radius can be converted into the primal problem of standard2-norm soft margin multiple kernel learning SVM. Finally, iterative optimization and SMO algorithm are used to find the optimal weight coefficients of basis kernel functions and solve the dual problem of multiple kernel learning SVM including the radius. Experiment results show that classifier based on2-norm soft margin multiple kernel learning SVM including the radius has a higher identification accuracy.