Discriminative Training Based Quadratic Classifier for Handwritten Character Recognition.pdf

September 7, 2007 17:48 WSPC/115-IJPRAI SPI-J068 00577 International Journal of Pattern Recognition and Artificial Intelligence Vol. 21, No. 6 (2007) 1035–1046 c? World Scientific Publishing Company DISCRIMINATIVE TRAINING BASED QUADRATIC CLASSIFIER FOR HANDWRITTEN CHARACTER RECOGNITION RUI ZHANG?, XIAO QING DING and HAI LONG LIU State Key Laboratory of Intelligent Technology and Systems Department of Electronic Engineering Tsinghua University, Beijing, P. R. China ?zhangrui96@tsinghua.org.cn In offline handwritten character recognition, the classifier with modified quadratic discriminant function (MQDF) has achieved good performance. The parameters of MQDF classifier are commonly estimated by the maximum likelihood (ML) estimator, which maximizes the within-class likelihood instead of directly minimizing the classifi- cation errors. To improve the performance of MQDF classifier, in this paper, the MQDF parameters are revised by discriminative training using a minimum classification error (MCE) criterion. The proposed algorithm is applied to recognizing handwritten numer- als and handwritten Chinese characters, the recognition rates obtained are among the highest that have ever been reported. Keywords: Pattern recognition; handwritten character; quadratic discriminant function; discriminative training. 1. Introduction In the statistical approaches to offline handwritten character recognition, the clas- sifiers are generally designed towards implementing the Bayesian decision. Given a pattern x which belongs to one of the N classes C1, C2, . . . , CN , if we have full knowledge of the prior probabilities p(Cn) and the class conditional densities p(x/Cn), the following Bayesian decision rule (1) will lead to a minimum classifi- cation error probability. C(x) = Ci, i = arg max n p(Cn)p(x/Cn), (1) where C(x) denotes a classification operation. In character recognition, the prior probabilities p(Cn) can usually be assumed equal, but neither the function form nor the para