期刊首页 优先出版 当期阅读 过刊浏览 作者中心 关于期刊 English

《中国工程科学》 >> 2009年 第11卷 第10期

一适用于超多类手写汉字识别的新改型Adaboost算法

清华大学电子工程系智能技术与系统国家重点实验室,北京 100084

收稿日期 :2009-08-24 发布日期 :2009-10-12 14:03:22.000

下一篇 上一篇

摘要

提出一种适用于超多类手写汉字识别的新改型Adaboost算法,采用基于描述性模型的多类分类器(modified quadratic discriminant function,MQDF)作为Adaboost基元分类器,可直接进行多类分类,无需将多类问题转化为多个两类问题处理,其训练复杂度大大低于已有的多类Adaboost算法。算法提出根据广义置信度更新样本权重,实验证明这种算法适用于大规模多类分类问题。为了降低算法的识别复杂度,提出从所有训练后得到的Adaboost基元分类器组中选择一个最优的基元分类器作为最终分类器的方法进行删减。在HCL2000及THOCR-HCD数据集上进行实验证明,所提改型Adaboost算法提高了识别率的有效性,该算法的相对错误率比现有最优算法分别下降了14.3 %,8.1 %和19.5 %。

图片

图1

图2

图3

图4

图5

图6

图7

参考文献

[1]  Friedman J, Hastie T, Tibshirani R.Additive logistic regression: a statistical view of boosting[ J] .The Annals of Statistics, 2000 , 28 ( 2 ) :337 -407

[2]  Schapire R E, Singer Y.Improved boosting algorithms using confi- dence -rated predictions [ J] .Machine Learning, 1999 , 37 ( 3 ) : 297 -336 链接1

[3]  Freund Y, Schapire R E.A decision -theoretic generalization of on -line learning and an application to boosting [ J] .Journal of Computer and System Sciences, 2006 , 55 ( 1 ) : 119 -139 链接1

[4]  Guruswami V, Sahai A.Multiclass learning, boosting, and error -correcting codes [ A] .Proceedings of the twelfth Annual Con- ference on Computational Learning Theory [ C ] . Santa Cruz, USA: ACM, 1999 :145 -155 链接1

[5]  Schapire R.Using output codes to boost multiclass learning prob- lems[ A] .Proceedings of the fourteenth International Conference on Machine Learning [ C ] .Nashville, USA: Morgan Kaufmann Publishers Inc, 1997 :313 -321 链接1

[6]  Liu C L, Fujisawa H.Classification and learning for character rec- ognition: comparison of methods and remaining problems [ A ] . Proceedings of the First IAPR TC3 Workshop on Neural Networks and Learning in Document Analysis and Recognition [ C] .Seoul, Korea: IEEE, 2005 :1 -7

[7]  Liu H L, Ding X Q.Handwritten character recognition using gra- dient feature and quadratic classifier with multiple discrimination schemes[ A] .Proceedings of the Eighth International Conference on Document Analysis and Recognition [ C ] . Seoul, Korea: IEEE, 2005 :19 -25 链接1

[8]  Liu H L, Ding X Q.Improve handwritten character recognition performance by heteroscedastic linear discriminant analysis [ A ] . The Eighteenth International Conference on Pattern Recognition [ C] .Hongkong, China, 2006 :880 -883 链接1

[9]  Kimura F, Takashina K, Tsuruoka S, etc.Modified quadratic dis- criminant functions and its application to Chinese character recog- nition [ J] .IEEE Trans on Pattern Analysis and Machine Intelli- gence, 1987 , 9 ( 1 ) : 149 -153 链接1

[10]  Lin X F, Ding X Q, Chen M, etc.Adaptive confidence trans- form based classifier combination for Chinese character recogni- tion [ J] .Pattern Recognition Letters, 1998 , 19 ( 10 ) : 975 - 988 链接1

相关研究