基于镜像学习和复合二次距离的手写汉字识别

为解决手写汉字识别中的相似字混淆问题,提出了一种基于镜像学习和复合二次距离的识别算法,提高现有的二次分类器对相似汉字的鉴别能力。该算法为识别置信度较低的训练样本生成镜像虚拟样本,通过迭代训练来调整易混淆字符类别间的分类界面,并对二次分类器给出的候选字使用复合二次距离进行两两鉴别,以减少识别错误。在HCL 2000样本库上的实验表明,该算法能有效提高手写汉字识别的性能,测试集上的误识率下降了20%。

Handwritten Chinese character recognition based on mirror image learning and the compound Mahalanobis function

In handwritten Chinese character recognition,many misclassification errors come from the confusion of similar characters.A recognition algorithm based on mirror image learning(MIL) and the compound Mahalanobis function(CMF) was developed to improve the discrimination ability of the quadratic classifier on the similar Chinese characters.The algorithm generates virtual image mirror samples from ambiguous training samples,and adjusts the classification surfaces between confused classes by an iterative learning scheme.The candidates given by the quadratic classifier are further discriminated using the compound Mahalanobis function,which efficiently reduces the classification errors.Tests on the HCL2000 database demonstrate that the algorithm effectively improves the recognition accuracy of handwritten Chinese character,reducing the misclassification rate for the test set by 20%.