Presentation 2004/2/13
String-Level Discriminative Training of Classifiers for Handwritten Numeral String Recognition
Cheng-Lin Liu, Katsumi Marukawa,
PDF Download Page PDF download Page Link
Abstract(in Japanese) (See Japanese page)
Abstract(in English) In handwritten numeral string recognition integrating segmentation and recognition, we have previously obtained superior results by training classifiers with segmented characters and non-character samples. This paper describes our works of string-level classifier training for further improving the string recognition performance. In this scheme, the classifier is initially trained on segmented characters, and then the classifier parameters are adjusted on string samples. During training, the string samples are dynamically segmented by candidate pattern classification and path search, and the classifier parameters are adjusted on segmented patterns under the minimum classification error (MCE) criterion of Juang et al. We tested the effectiveness of string-level training with various classifier structures on the numeral string images of NIST Special Database 19. It was observed that string-level training yields higher string recognition accuracy than character-level training via decreasing segmentation errors while sacrificing the classification accuracy on segmented characters. By combining the string-level trained classifier and the character-level trained classifier, we have achieved even higher string recognition accuracy.
Keyword(in Japanese) (See Japanese page)
Keyword(in English) Numeral string recognition / integrated segmentation and recognition, / character classification / string-level training / minimum classification error (MCE) criterion / classifier combination
Paper # TL2003-46,PRMU2003-232
Date of Issue

Conference Information
Committee PRMU
Conference Date 2004/2/13(1days)
Place (in Japanese) (See Japanese page)
Place (in English)
Topics (in Japanese) (See Japanese page)
Topics (in English)
Chair
Vice Chair
Secretary
Assistant

Paper Information
Registration To Pattern Recognition and Media Understanding (PRMU)
Language ENG
Title (in Japanese) (See Japanese page)
Sub Title (in Japanese) (See Japanese page)
Title (in English) String-Level Discriminative Training of Classifiers for Handwritten Numeral String Recognition
Sub Title (in English)
Keyword(1) Numeral string recognition
Keyword(2) integrated segmentation and recognition,
Keyword(3) character classification
Keyword(4) string-level training
Keyword(5) minimum classification error (MCE) criterion
Keyword(6) classifier combination
1st Author's Name Cheng-Lin Liu
1st Author's Affiliation Central Research Laboratory, Hitachi, Ltd.()
2nd Author's Name Katsumi Marukawa
2nd Author's Affiliation Central Research Laboratory, Hitachi, Ltd.
Date 2004/2/13
Paper # TL2003-46,PRMU2003-232
Volume (vol) vol.103
Number (no) 659
Page pp.pp.-
#Pages 6
Date of Issue