基于并行特征组合与广义K-L变换的字符识别

针对传统的串行特征融合方法的弱点,提出了一种新的并行特征融合方法.该方法的基本思路是:首先,利用复向量将样本空间上的两组特征集组合起来,构成复特征向量空间;然后,从理论上推广了经典的K-L变换方法与3种基本的K-L展开方法,使其适用于复特征向量空间内的特征抽取.此外,还揭示了并行特征融合的对称性质,并详细讨论了并行特征组合的策略问题.最后,用所提出的方法来解决手写体字符的特征抽取与识别问题.在南京理工大学NUST603HW手写体汉字库以及Concordia大学的CENPARMI手写体阿拉伯数字数据库上的实验结果表明,所提出的特征融合方法不仅较大幅度地提高了识别率,而且识别结果优于传统的串行特征融合方法.

Handwritten Character Recognition Based on Parallel Feature Combination and Generalized K-L Expansion

Considering the weaknesses of traditional serial feature fusion technique, a novel parallel features fusion method is proposed in this paper. The main idea of this method can be described as follows. First of all, two sets of feature vectors corresponding to a same sample space are combined together via complex vectors, which are used to construct a complex feature vector space. Then, the classical K-L transform and K-L expansion methods are developed theoretically to suit for feature extraction in the complex feature space. Moreover, the symmetric property of parallel feature fusion is revealed, and, how to combine features effectively is discussed in detail. Finally, the proposed method is used to solve the handwritten character feature extraction and recognition problems. Experiments are performed on NUST603 handwritten Chinese character database built in Nanjing University of Science and Technology as well as the well-known CENPARMI handwritten digit database of Concordia University. The experimental results indicate that the recognition rates are improved significantly after parallel feature fusion, and the proposed parallel features fusion method is superior to the traditional serial feature fusion one.