Chinese Journal of Computers   Full Text
Title	Piecewise Support Vector Machines
Authors	REN Shuang-Qiao YANG De-Gui LI Xiang ZHUANG Zhao-Wen
Address	(School of Electronic Science and Engineering, National University of Defense Technology, Changsha 410073)
Year	2009
Issue	No.1(77—85)
Abstract &   Background	Abstract A novel piecewise support vector machines (PSVM) model is provided in this paper, which used the traditional piecewise linear recognition method for reference. In this new model, the feature space was firstly partitioned into several subspaces, then the piecewise classification surface was developed by linking the optimal classification surfaces in each subspaces based on SVM. As the experimental results shown, validated with the Two Spiral Data and the actually measuring data, the performance of PSVM such as computational efficiency, classification capability and generalized capability are improved obviously in contrast to SVM. Keywords piecewise linear; SVM; space partition; piecewise; recognition Background In recent years, many scholars have carefully studied the linear classifier because of little complex form and computational time. When the samples can not be correctly classified by the linear classification hyper plane, the linear classifier will be not competent for recognition the samples. In the interesting of correctly recognizing the samples with complex distribution, Vladimir N.Vapnik introduced the kernel function technique and support vector machines. The kernel functions in SVM allow the classifier to be extremely flexible to suit the requirements of the classification problem, but it is an on-going issue that the kernel function can be adaptively selected and constructed. Additional, the nonlinear SVM essentially is a linear classifier in the high dimension feature space which is educed through the nonlinear kernel function. When the samples complexly distributing, the performance of nonlinear SVM, such as the classification capability, generalized capability and the computation time, will encounter the hard challenge. In classical pattern classification theory and methods, the piecewise linear recognition technique is a compromise between linear and nonlinear recognition method. Its classification surface is a special nonlinear function, which consists of several classification hyper planes. Comparing to the general nonlinear classification hyper surface, the piecewise linear classification surface is more simple and has the ability to approach a great variety of optimal classification nonlinear hyper surface. A novel piecewise support vector machine (PSVM) model based on piecewise linear classification concept is provided in this paper. In this model, the feature space was firstly partitioned into several subspaces, and the optimal classification surface was constructed based on SVM in each subspace, then the piecewise classification surface in feature space was developed by linking the optimal classification surfaces in subspaces. As the experimental results shown, the performance of PSVM such as computational efficiency, classification capability and generalized capability are improved obviously in contrast to classical SVM. This work is supported by the National Science Foundation of China, under grant No.60402032. The aim of this project is to solve the automatic recognition for radar observing objects. This paper offeres the kernel function adaptively construction method, which is an important portion to the automatic objection recognition using SVM.