Fast self-generation voting for handwritten Chinese character recognition
- PDF / 1,125,126 Bytes
- 12 Pages / 595.276 x 790.866 pts Page_size
- 20 Downloads / 215 Views
ORIGINAL PAPER
Fast self-generation voting for handwritten Chinese character recognition Yunxue Shao · Chunheng Wang · Baihua Xiao
Received: 20 May 2012 / Revised: 7 October 2012 / Accepted: 11 October 2012 / Published online: 27 October 2012 © Springer-Verlag Berlin Heidelberg 2012
Abstract In this paper, a fast self-generation voting method is proposed for further improving the performance in handwritten Chinese character recognition. In this method, firstly, a set of samples are generated by the proposed fast selfgeneration method, and then these samples are classified by the baseline classifier, and the final recognition result is determined by voting from these classification results. Two methods that are normalization-cooperated feature extraction strategy and an approximated line density are used for speeding up the self-generation method. We evaluate the proposed method on the CASIA and CASIA-HWDB1.1 databases. High recognition rate of 98.84 % on the CASIA database and 91.17 % on the CASIA-HWDB1.1 database are obtained. These results demonstrate that the proposed method outperforms the state-of-the-art methods and is useful for practical applications. Keywords Handwritten Chinese character recognition · Fast self-generation voting · Line density equalization · Normalization-cooperated feature extraction · Modified quadratic discriminant function
1 Introduction The problem of handwritten Chinese character recognition (HCCR) has been investigated over a long time for its Y. Shao (B) · C. Wang · B. Xiao Institute of Automation Chinese Academy of Sciences, 95 Zhongguancun East Road, 100190 Beijing, China e-mail: [email protected] C. Wang e-mail: [email protected] B. Xiao e-mail: [email protected]
potential in many applications. Many methods have been proposed and very high recognition rate has been obtained on most of the databases. In the character normalization stage, many methods have been proposed to reduce the within-class shape variation, including the nonlinear normalization methods [1–3], the bi-moment method [4] and the pseudo 2D normalization method [5]. In feature extraction stage, a large variety of feature extraction methods are proposed [6]. The stroke direction feature which can be measured from skeleton [7], chain-code [8] or gradient [9–11] is the most popularly used due to the high performance and the ease of implementation. Previous studies have shown that the gradient feature [11] outperforms the other stroke direction features and the Gabor filter feature [12]. After character normalization and feature extraction, classifiers based on quadratic discriminant functions (QDFs) are usually applied to HCCR. The most popular one is the modified quadratic discriminant functions (MQDF1 and MQDF2) proposed by Kimura et al. [8]. The MQDF improves the generalization performance via replacing the eigenvalues in the minor subspace of each class with a constant. Besides MQDF1 and MQDF2, the pseudo Bayes classifier [13], asymmetric Mahalanobis distance (MD) [14], determinant normalized QDF [
Data Loading...
