In multi-class recognition, neighbor classes in the original feature space are prone to be more confused after feature dimensionality reduction by linear discriminant analysis ( LDA). It does not benefit the recognition. To solve this problem, this paper proposes a similar pattern discriminant analysis ( SPDA) method, which constructs the feature transformation matrix based on discriminant vectors extracted from similar pattern pairs. The proposed SPDA method was applied together with LDA to the cascade modified quadratic discriminant function ( MQDF) classifiers to improve the performance of recognizing handwritten Chinese characters. The results show that the recognition accuracy on handwritten character library 2000 ( HCL2000) reaches up to 98. 82% , which is higher than the corresponding results found in the literature. The experiment indicates that the proposed method is effective.%针对多类识别时原始特征空间中相近的类经过线性鉴别分析(LDA)降维后,在低维空间中易被混淆,不利于识别的问题,提出了一种通过对相似类对抽取鉴别向量构成特征变换矩阵的相似模式类鉴别分析(SPDA)方法,并将该方法与LDA降维相结合,应用于级联改进二次鉴别函数(MQDF)分类器中,实现了对手写汉字识别性能的进一步提高.在脱机手写汉字字符集2000(HCL2000)上的识别率为98.82％,识别结果高于可查文献中相应的识别结果,这表明该方法是有效的.
展开▼
