о뷢չ
JOURNAL OF COMPUTER RESEARCH
AND DEVELOPMENT
1999 36 12 Vol.36 No.12 1999



дЧĳȡʶ
ҡɽˡﾸ
ժҪ˻ںʹƵĶʶĹƷ˻ھͳƲԵѼ任KL任ȡдֵЧķ.ʵConcordia University CENPARMIдݿ.ھطдֵ12ʶʵ.ʵƷڳõͶƱƷַƷʶʸߴ97%.һϸĽṹƷһɷüɷ˸õʵʶʡʶɿԷֱɴﵽ97.15%2.05%99.18%Ŀǰڸдݿõõʵ.
ؼʡģʽʶȡдʶ𣬶
ͼšTP391
EXTRACTION AND RECOGNITION OF EFFECTIVE
DISCRIMINANT FEATURES FOR HANDWRITTEN DIGITS
JIN Zhong, HU Zhong-Shan, YANG Jing-Yu, LIU Ke*, and SUN C. Y.*
(Department of Computer Science, Nanjing University of Science
& Technology, Nanjing 210014)
(Center of Pattern Recognition and Machine Intelligence,
Concordia University, Montreal)
AbstractPresented in this paper is an estimating method of multi-feature and multi-classifier combination based on the posterior probability estimators. Also presented is a method to extract effective discriminant features of handwritten digits based on a set of uncorrelated optimal discriminant features and KL transform. Experiments have been performed with Concordia University CENPARMI's handwritten digit database based on the nearest-neighbor distance classifier and the nearest-neighbor correlation classifier, and 12 features of handwritten digits. Experimental results show that the estimating method is better than the polling method or the counting method respectively and the recognition rate of the estimating method is as high as 97%. A new combination classifier is finally brought forward, which is based on the strict structure classifier and the estimating method. Better experimental results have been obtained by means of this new combination classifier: the recognition rate, the reject rate, and the reliability are as high as 97.15%, 2.05%, and 99.18% respectively, which are the best results up to now on the handwritten digit database.
Key wordspattern recognition, feature extraction, discriminant analysis, handwritten digit recognition, multi-classifier combination
1
дʶǼģʽʶһǰؿ.дʶоУѶֵͼȡķ֣1,2ݣõͳGabor任3ݡLegendre4ݡPseudo-Zernike5Zernike6.ͳͨӳͼֲЩͼԤ򵥣УֵһЩϸṹַӦ.ͼĹǼܺɳȡдֵĽṹֽṹϸַӳͼԤӣУһƵĻڽṹķʮ.дʶоУõķڷ7ԼϷ8,9.һΪͬʵӳĲͬ棬һռֵģʽһռϺ׷ֿӦͬһĲִͬӲͬĽǶȽӳ䵽߿ռϣòͬͲͬϾͿȫ淴ӳһ壬ӶõһϺõķ.һںʹƵĶϵĹƷ˻Ѽ任KL任ɳõͳȡдֵЧķһϸĽṹ9Ʒһɷ.
2ϾߵĹƷ
2.1Bayesо
cģʽ𣺦12ciģʽѵniΪXis(s=1,2,,ni).XΪһBayesоݺP(i|X)ԭXࣺ

(1)
2.2ʵĹ
K.ÿķߺֵɸP(i|X)ĹƺʹӦǸһ.õھطĺʹƷĺʹƷɲο.
ھǰŷϾС.XÿѵŷϾdis

(2)
ʵһƹʽ

(3)
طǰϵ.Xÿѵϵis

(4)
ʵһƹʽ

(5)
2.3ϾߵĹƷ
ģʽXɶX1,X2,,XQɣɷֱʹKʶ𣬸ʵĹƽֵ

(6)
ǺP(i|X)ƫƣɸеմ֪ģʽĿܴĿܶʱƽֵںP(i|X).Ķھطĺʹƹʽܱ֤ΪP(i|X)ƫ.
ϾߵĹƷںʹƣ(6)ʽƽֵԭ.
õĶϷͶƱƷַ.ͶƱõӶԭ򣬼֧ĳһߵķΪһ߾յĽ.Ʒַǽÿһڸеӣ11֣kk֣ӶõںϺĵ÷ܺͣȡܷСΪƷַѡ.
ȻͶƱƷַ˵ΪʵĹȡΪʽĹƷ.ںʹƵĹƷһܸõøľߺֵṩϢ.
3ͳƲԵѼ任
ģʽXΪNάʸģʽcģʽɢΪSwɢΪSbɢΪSt.Fisher׼£

(7)
ЦΪNά.
ʹFisher׼(7)ʽȡֵ1FisherѼģʽڦ1ϵͶӰʹģʽСɢɢ.1йֵӦ

(8)
ͳƲԵѼ任10ݾrr1Ѽ1,2,,r󣬵r+1Ѽ㹲(9)ʽµʹFisher׼(7)ʽȡֵr+1

(9)
1.ͳƲԵѼ任ĵr+1Ѽr+1йֵӦ

(10)
ʽУ

(11)
IΪλ󣬶

(12)
֤ף10.
4дֵȡ
ĲõݿǹϹ㷺ʹõConcordia University CENPARMIдݿ⣬4000ѵ2000.
4.1ͼ
ף9ͨдͼԤȡ4ͼ4ͼݴСΪ20.2MB
XG256άGabor任3ݣ
XL121άLegendre4ݣ
XP36άPseudo-Zernike5ݣ
XZ30άZernike6.
4.2Чȡ
Fukunagaά11ΪԺcģʽķ⣬ЧѼʸc-1.дֵΪ10ЧļΪ9.дֵ4ͼXGXLXPXZֱͳƲԵѼ任õ4ЧļֱΪXG1XL1XP1XZ1õдֵ4
XG1XGѼ任ȡ9ά
XL1XLѼ任ȡ9ά
XP1XPѼ任ȡ9ά
XZ1XZѼ任ȡ9ά.
4.3KL任Чȡ
ȻģʽĿ϶ʱʹ(6)ʽһܶԺʹƵýϺãϣдֳȡ.ڱдֵ4ͼXGXLXPXZȷֱԸԵؾRbΪ12KL任.ؾRbΪ10Ӷɷֱõ410άдͼѹȻٶ410άѹͳƲԵѼ任õ4µ9άֱΪXG2XL2XP2XZ2ֵõдֵ4
XG2XGKL任Ѽ任ȡ9ά
XL2XLKL任Ѽ任ȡ9ά
XP2XPKL任Ѽ任ȡ9ά
XZ2XZKL任Ѽ任ȡ9ά.
5дʶʵ
5.1ʶ
Matlab̼дֵ8Чļھطֱдֵ12ʶʵ飬ʶʵʵ1ϰ벿.ͶƱƷַƷֱ122ڷϣʶʵʵ1°벿.ͶƱƷַߵ÷ͬʱʶ.ԹƷںʵֵͬʱʶ.
1дʶʵ
ھط
Gabor任XG0.17850.1440
XG10.17500.2120
XG20.21300.2585
LegenderXL0.08750.0665
XL10.10750.1405
XL20.12900.1385
Pseudo-ZernikeXP0.23650.2410
XP10.29850.3870
XP20.35400.4360
ZernikeXZ0.24450.2540
XZ10.30750.3840
XZ20.37750.4550
ʶͶƱ(ʶ)
Ƿַ(ʶ)
Ʒ

ӱ1Կܵʶʽϴʶ𷽷ήʶʣһںʹƵĹƷͶƱƷַãʵУƷδʶʹͶƱƷַľʶʣƷʶֻ3%ҲֱͶƱʶƷַʶ.һʵеڸдݿõʵ΢һЩҲһܺõʵ.
5.2ɷʶʵ
ף9һϸĽṹֵĹǼܻõĽṹʹһЩǳ򵥵һͺܿɿرʶڷ׼Ϊֻв׼г£ͼĹǼûж˵㡢ûӵļ򵥱պȦʱͼ0ͼĹǼ˵ļֱʱͼ1ͼĹǼ˵㡢һϷ۵㣬۵˵һȵֱʱͼ7ȵ.ýṹʶΪ54.95%ʶȻʶʺܸߣɿҲܸߣﵽ100%.ͼ1ʶӣͼ2Ǿʶ. 

ͼ1ʶ

ͼ2ʶ
Ϊ˽һдֵʶʵڻṹʶĹƷһɷýṹʶ࣬ԱṹʶƷ122ʶ𣬶ҵʹƵĴδֵʹƵֵֵֵ֮ʱʶ

(13)
ʱʶ.ɷʶʵ2.
2ɷʶʵ
ֵ0.9940.9951ף7ף8ף9
ʶ0.02050.0155000.010.0205
ʶ(41)(31)
0.0080.01050.01850.0290.0160.009
(16)(21)(37)
ʶ0.97150.9740.98150.9710.9740.9705
ɿ0.99180.98930.98150.9710.98360.9908

ӱ2ԿǾʶڦ=1ʱļɷʶʵף7ݵʶʵҪϸߵĿɿϸߵʶʣڦ=0.995ʱļɷʶʵף8ݵʶʵҪʶɿԴﵽ99%ϣڦ=0.994ʱļɷʶʵף9ݵʶʵ. 
6ᡡ
˻ںʹƵĶʶĹƷ˻ھͳƲԵѼ任KL任ȡдЧķ.ʵĹƷڶʶõͶƱƷַùƷ122ڷʶʿɴﵽ97%.һϸĽṹ9ƷһɷüɷʶʡʶɿԷֱɴﵽ97.15%2.05%99.18%һĿǰڸдݿõʵ.
õȻѧ(Ŀ69672013)͹ʺоĿ.
߼飺ңУ196112ڣʿоҪоΪģʽʶд
ʶʶ.
ɽУ19732ʿоҪоΪģʽʶдʶ
.
У194112ʿʦڣҪоΪӾϢ
ںϡģʽʶܻ.
,,1960,,Concordiaѧַָʶо.
ﾸ,Concordiaѧ,CENPARMIо,ַʶѧ.
ߵλҡɽ(ϾѧѧϵϾ210014)
˪ﾸ(Concordiaѧģʽʶġ)
ο
1Shunji M, Suen C Y, Kazuhiko K. Historical review of OCR research. Proc of IEEE, 1992, 80(7): 10291058
2Oivind T, Anil J K, Torfinn T. Feature extraction methods for character recognitionA Survey. Pattern Recognition, 1996, 29(4): 641662
3Yoshihiko H et al. Recognition of handwritten numerals using Gabor features. In: Proc of the 13th ICPR, Austria, 1996. 250253
4Liao S X, Pawlak M. On image analysis by moments. IEEE Trans Pattern Anal Machine Int' l, 1996, 18(3): 254266
5Bailey R R, Mandyam S. Orthogonal moment feature for use with parametric and non-parametric classifiers. IEEE Trans Pattern Anal Machine Intell, 1996, 18(4): 389398
6Alireza K, Yawhua H, Invariant image recognition by Zernike moments. IEEE Trans Pattern Anal Machine Int' l, 1990, 12(5): 489497
7Lee S W. Off-line recognition of totally unconstrained handwritten numerals using multilayer cluster neural network. IEEE Trans Pattern Anal Machine Intell, 1996, 18(6): 648652
8Gader P D. Automatic feature generation for handwritten digit recognition. IEEE Trans Pattern Anal Machine Int' l, 1996, 18(12): 12561261
9ɽ,¦,,,ﾸ. ڶϵдʶ.ѧ,1999, 22(4): 369374
(Hu Zhongshan, Lou Zhen, Yang Jingyu, Liu Ke, Suen C Y. Handwritten digit recognition based on multi-classifier combination. Chinese Journal of Computers(in Chinese), 1999, 22(4): 369374)
10,,½. һ־ͳƲԵѼʸ. ѧ,1999, 22(10): 11051108
(Jin Zhong, Yang Jingyu, Lu Jianfeng. An optimal set of uncorrelated discriminant features. Chinese Journal of Computers(in Chinese), 1999, 22(10): 11051108)
11Fukunaga K. Introduction to Statistical Pattern Recognition. New York: Academic Press, 1990
12ңɽ. BPʶ𷽷. о뷢չ, 1999, 36(3): 274277
(Jin Zhong, Hu Zhongshan, Yang Jingyu. A face recognition method based on the BP neural network, Journal of Computer Research and Development(in Chinese), 1999, 36(3): 274277)
ԭյڣ1999-01-22޸ĸյڣ1999-05-10.
