Adaptive Support Vector Machine And Its Application In Handwritten Chinese Character Recognition

In the early 90's, a new learning method had been proposed based on the statistical learning theory(SLT), called support vector machine(SVM), which is a small-sample statistics and concerns mainly the statistic principles when samples are limited. SVM is is a research hotspot at present in the domain of Pattern Recognition (PR) and Regression Analysis (RA).There are mainly two focuses of SVM including support vector classification (SVC) and support vector regression (SVR), while the research of SVR is not a patch on SVC either in theory or in applications. This thesis gives a detailed review of the SVM, proposed a new as well as an improved algorithm based on the adaptron problem. The main research contents of this thesis include two parts: the theory part and the application part.1 .The theory part of this thesis also includes two parts:⑴The first part gives a comprehensive survey of statistical learning theory and binarySVM.Then it gives a detailed discussion for some related knowledge like kernel methods and optimization problems.⑵The second part has proposed a new algorithm-AWSVM. In the weighted support vector machines for regression, each training sample had different approximation error requirement and different penalty due to the effect of weighting factors on them. In order to solve the shortcoming of the weighting factors selection problems in WSVR, an adaptive selecting approach is proposed which can choose appropriate weighting factors adaptively by the new regression algorithm. Experimental results show that the proposed method has a better performance.2.In the application part, The Handwritten Chinese Character Recognition (HCCR) is analyzed. An improved algorithm is presented for classification in high dimensional spaces based on Adatron algorithm in Gaussian Radial Basis Kernel Function. The adjusted zi , detailed discussedαiand limited number of iterations are considered in the new algorithm in order to make adatron-kernel suited for HCCR. Our experiment results show that the performance is improved with the proposed algorithm both in recognition and Kernel parameter chosen.