ARTICLE IN PRESS. Neurocomputing

Size: px
Start display at page:

Download "ARTICLE IN PRESS. Neurocomputing"

Transcription

1 Neurocomputng 73 (2010) Contents lsts avalable at ScenceDrect Neurocomputng journal homepage: Incremental learnng of LDA model for Chnese wrter adaptaton Lanwen Jn, Ka Dng, Zhbn Huang School of Electronc and Informaton Engneerng, South Chna Unversty of Technology, Guangzhou, Guangdong , Chna artcle nfo Avalable onlne 12 March 2010 Keywords: Incremental LDA WILDA Wrter adaptaton Handwrtten Chnese character recognton abstract A new wrter adapton method based on ncremental lnear dscrmnant analyss (ILDA) s presented n ths paper. We frst provde a more general soluton for ILDA and then present a Weghted ILDA (WILDA) approach. Based on ILDA or WILDA, the wrter adaptaton s performed by updatng the LDA transformaton matrx and the classfer prototypes n the dscrmnatve feature space. Expermental results show that both ILDA and WILDA are very effectve to mprove the recognton accuracy for wrter adaptaton, and WILDA outperforms ILDA. The proposed WILDA based wrter adaptaton method can reduce as much as 47.88% error rate on the wrter-dependent dataset whle t only has as less as 0.85% accuracy loss on the wrter-ndependent dataset. It ndcates that wrter adapton usng WILDA can sgnfcantly ncrease the recognton accuracy for the partcular wrter whle havng lmted mpact on the accuracy for general wrters. & 2010 Elsever B.V. All rghts reserved. 1. Introducton The ablty to transcrbe handwrtten characters to a computerzed text format s a great beneft to nputtng, organzng and annotatng data n varous applcatons such as the nput, storage and dstrbuton of notes or messages [3,22,32]. The successes of products, such as PDA, smart cell phone, and Tablet PC, are the evdence that users have nterest n such capabltes. However, the large varablty of handwrtng styles across ndvduals makes handwrtng recognton a challengng problem. Although great progress has been acheved n the feld of onlne handwrtten Chnese character recognton (OHCCR) durng the past 40 years [8,12,22,25,32], recent researches on unconstraned cursve onlne handwrtng recognton show that ths problem s far from havng been completely solved. Jn et al. reported that for the recognton of 6763 categores of unconstraned handwrtten characters from the SCUT-COUCH dataset [21], the best recognton accuracy was only 92.43% usng the state-of-the-art feature extracton method plus lnear dscrmnant analyss (LDA) classfer [5]. Wang et al. [39] reported that usng a state-of-theart recognzer for the classfcaton of the samples from a new unconstraned onlne handwrtten dataset CASIA-OLHWDB1, the hghest accuraces acheved on average, 92.44% for 4037 categores and 92.91% for 3866 categores, respectvely, were acheved, whch are much lower than those reported on other prevous databases, where the handwrtten samples were much more regularly wrtten (e.g., 98.56% on HCL2000 [24], 97.84% and Correspondng author. E-mal address: lanwen.jn@gmal.com (L. Jn) % on Japanese Kanj [23]). On the other hand, the requred recognton rate for the recognzer by ordnary users s very hgh. For example, tests wth keyboard typng have shown that the wrters tolerate random errors up to 1% whle 0.5% s unnotceable and 2% s ntolerable [9]. All of these ndcate that the unconstraned OHCCR problem s far from beng completely solved, especally when usng some recently avalable challengng datasets, such as SCUT-COUCH [21] or CASIA-OLHWDB1 [39]. One challenge of unconstraned OHCCR we face s that there are too many dfferent wrtng styles to be handled when desgnng a general purpose handwrtng recognzer. Wrter-ndependent systems traned from examples requre large tranng sets from many wrters to deal wth ths varablty, but t cannot yet acheve very hgh performance for unconstraned OHCCR [21,39]. In contrast, wrter-dependent systems can be traned on a specfc user s handwrtten samples to acheve hgher accuracy [33]. It s generally agreed that, for a gven handwrtng recognton task, a wrter-dependent system usually outperforms a wrter-ndependent system [38]. The wrter adapton s the process of convertng a wrter-ndependent system learned from the wrter-ndependent dataset to a wrter-dependent system, whch s turned for a partcular wrter usng a specfc ncremental data. As the wrteradaptaton s an onlne ncremental learnng process to learn the partcular wrtng behavor and adaptvely updatng the classfcaton model, we usually need an ntal classfcaton model traned on general-purpose wrter-ndependent dataset and then conduct the adaptaton learnng processng. Ths adapton has the potental advantage of sgnfcantly ncreasng recognton accuraces for a partcular wrter, whch s very useful for a real world applcaton, such as buldng a hgh performance, wrter adaptve (personalzed) onlne handwrtten character nput /$ - see front matter & 2010 Elsever B.V. All rghts reserved. do: /j.neucom

2 L. Jn et al. / Neurocomputng 73 (2010) method. In the past, a number of wrter adaptaton handwrtng recognton methods have been proposed [3,7,16,20,33,38]. Szummer and Bshop [33] proposed a dscrmnatve wrter adaptaton method through clusterng the wrtng styles, tranng a set of correspondng classfers and then choosng an approprate combnaton of classfers for a partcular wrter. Connel and Jan [3] proposed an adaptve onlne handwrtng recognton model, where another wrter-dependent model was used to dentfy the styles present n a partcular wrter s tranng data, and then these models are retraned usng the wrter s data. But the adaptaton n ths way depends on the correctly clusterng users wrtng styles by classfcaton confdence. Vuor [38] proposed a smple prototype based adaptaton system usng k nearest neghbor (KNN) classfer. The whole adaptaton conssted of three steps,.e. addng new prototypes, deactvatng confusng prototypes, and reshapng exstng prototypes. Kenzle and Chellaplla [16] presented a personalzed handwrtng recognton approach by mnmzng a regularzed rsk functon of SVM. LaVola and Zeleznk [20] proposed a practcal technque of usng a wrter-ndependent recognzer to mprove the accuracy of wrter-dependent symbol recognzer based on the AdaBoost learnng algorthm. Unfortunately, all these methods are desgned for small scale handwrtng recognton problem (for example, Englsh letter, dgt or symbol recognton), where the class number s relatvely small; thus many of the adaptaton methods are not practcally applcable (such as SVM or AdaBoost based adaptaton methods) for handlng large datasets wth many classes, such as Chnese handwrtng recognton problem nvolvng thousands of classes and hundreds of thousands of tranng/ testng handwrtten samples. On the other hand, as a well known scheme for feature extracton and dmenson reducton, lnear dscrmnant analyss (LDA), also known as Fsher dscrmnant analyss (FDA), has been wdely used n OLHCCR [5,22,24,25] and other pattern classfcaton tasks [10,13,15,31]. The LDA seeks the best lnear projectons of data for dscrmnaton, under the assumpton that the classes have equal covarance Gaussan structure [6]. However, recent researches demonstrate that the classcal LDA has some problems [34]. The frst s heteroscedastc problem [27] that s the LDA models dfferent classes wth dentcal covarance matrces. Therefore, t fals to take account of any varatons n the covarance matrces between dfferent classes. The second s multmodal problem [11] that s the samples n each class cannot be approxmated by a sngle Gaussan n many applcatons. Instead, a Gaussan mxture mode (GMM) [2] s requred. However, the LDA models each class by a sngle Gaussan dstrbuton. The thrd s class separaton problem [35]. In applcatons, dstances between dfferent classes are dfferent and the LDA tends to merge classes that are close together n the orgnal feature space. The fourth s the small sample sze (SSS) problem [37,40],.e. the number of tranng samples s less than the dmenson of the feature space, whch s also known as sngularty problem [18]. To solve the heteroscedastc problem, Decell and Mayekar [4] proposed a method to obtan a subspace to maxmze the average nterclass dvergence, whch measured the separatons between the classes. Ths crteron takes nto account the dscrmnatve nformaton preserved n the covarances of dfferent classes. Kumar and Andreou [19] developed the heteroscedastc dscrmnant analyss (HDA) by droppng the dentcal class covarances assumpton. Jelnek [14] proposed a dfferent way to deal wth the heteroscedastc problem n subspace selecton by the gradent steepest ascent method to fnd the projecton matrx. Loog and Dun [27] ntroduced the Chernoff crteron to heteroscedastcze LDA. A straghtforward approach to solvng the mult-model problem s to employ the GMM approach. Haste and Tbshran [11] combned GMM wth LDA by drectly replacng the orgnal sngle Gaussan n each class by a Gaussan mxture model. To deal wth the class separaton problem, Lotlkar and Kothar [28] developed the fractonal-step LDA (FS-LDA) by ntroducng a weghted functon. Loog et al. [26] developed another weghted method for LDA, namely the approxmate parwse accuracy crteron (apac). The advantage of apac s that the projecton matrx can be obtaned by the egenvalue decomposton. Lu et al. [29] combned the FS-LDA and the drect LDA for very hgh dmensonal problems. However, both FS-LDA and apac do not use the dscrmnatve nformaton n dfferent class covarances. To reduce ths problem, Tao et al. [36] proposed the general averaged dvergences analyss framework by usng geometrc mean for subspace selecton. To deal wth the SSS problem, many approaches have been proposed, such as pseudo-nverse LDA, PCA+LDA and regularzed LDA [18]. In recent years, Ye et al. [40] proposed the two dmensonal LDA (2DLDA). Motvated by the successes of the 2DLDA, Tao et al. [37] proposed the general tensor dscrmnant analyss (GTDA) to solve the SSS problems. Although the LDA and ts extensons have been wdely used n pattern recognton feld, the typcal mplementaton of these technques assumes that a complete dataset for tranng s gven n advance and t s often benefcal to learn the LDA model from large tranng sets, whch may not be avalable ntally. Ths motvates technques for ncrementally updatng the LDA model when more data become avalable [17,30]. Several ncremental versons of LDA (ILDA) have been suggested [17,30,41], whch have successfully been appled to onlne learnng tasks such as the classfcaton of data streams [30], face mage retreval [17] and face recognton [41]. Although a number of researches on wrter adaptaton or ILDA were conducted, the ILDA based wrter adaptaton handwrtng recognton remans unexploted. Motvated by ths problem, we nvestgate how to adapt a wrter ndependent recognzer to make t wrter dependent based on the ncremental learnng of LDA model under the LDA based OLHCCR classfcaton framework for the frst tme n ths paper. We frst provde a general ncremental learnng soluton for LDA, and then propose a weghted ncremental lnear dscrmnant analyss (WILDA) approach for wrter adaptve handwrtng recognton by takng nto account the ssue of uncertan number of ncremental data for wrter adapton n an onlne handwrtng recognton applcaton. Based on the ncremental learnng of the LDA model usng ILDA or WILDA, the wrter adaptaton s performed by updatng the LDA transformaton matrx and the classfer prototypes n the dscrmnatve feature space. Expermental results show that both ILDA and WILDA are very effectve to mprove the recognton accuracy for partcular wrters, and WILDA outperforms ILDA. The expermental results ndcate that the wrter adapton usng the WILDA approach can not only sgnfcantly ncrease the recognton accuracy for the partcular wrters but also have lmted mpact on the accuracy for the general wrters. The rest of ths paper s organzed as follows. Secton 2 presents a general soluton for the ILDA algorthm, and then proposes a new weghted ncremental lnear dscrmnant analyss (WILDA) approach. Classfer desgn based on LDA and the wrter adaptaton based on ILDA/WILDA are gven n Secton 3. Secton 4 presents the expermental results and dscusson. Fnally, the conclusons are summarzed n Secton Incremental learnng of lnear dscrmnant analyss (LDA) model Pang et al. [30] proposed an ncremental LDA for classfcaton of data streams. However, the fnal soluton of ths method s too

3 1616 L. Jn et al. / Neurocomputng 73 (2010) complex, snce the sequental ncremental learnng condton and the chunk ncremental learnng condton are consdered separately. In addton, for each case, the soluton s dvded nto two cases dependng on whether the new class sample s added or not. In ths secton, we frst gve a bref ntroducton to LDA, and then present a general ncremental learnng soluton for LDA based on the research n [30]. Furthermore, consderng the problem of uncertan number of ncremental data for wrter adapton n an onlne handwrtng recognton applcaton, we present a Weghted ILDA (WILDA) method n Secton 2.3. ncremental samples, respectvely, whch are computed by m y ¼ 1 l X l m y ¼ 1 L j ¼ 1 X L ¼ 1 y ðþ j y ¼ 1 L X P ¼ 1 l m y The ncremental wthn-class scatter matrx S y w and between-class scatter matrx S y of the ncremental samples are gven as follows: b ð5þ ð6þ 2.1. Introducton to LDA X l ðy ðþ m y j ÞðyðÞ m y j ÞT ¼ 1 ¼ 1 j ¼ 1 S y w ¼ XP S y ¼ XP ð7þ LDA [6] s a supervsed learnng method, whch utlzes the category nformaton assocated wth each sample. The goal of LDA s to seek drectons for effcent dscrmnaton through maxmzng the between-class scatter whle mnmzng the wthn-class scatter. Mathematcally speakng, the wthn-class scatter matrx S w and between-class scatter matrx S b are defned as S w ¼ XM j ¼ 1 S b ¼ XM j ¼ 1 S j ¼ XM X N j ðx ðjþ j ¼ 1 ¼ 1 N j ðm j mþðm j mþ T m j Þðx ðjþ m j Þ T ð1þ where x ðjþ s the th sample of class j, m j ¼ð1=N j Þ P N j ¼ 1 xðjþ s the mean of class C j, m ¼ð1=NÞ P M j ¼ 1 N jm j s the mean vector of all classes, M s the number of classes, N j s the number of samples of class j, and N ¼ P M j ¼ 1 N j s the total sample number. For LDA transformaton matrx, W lda can be derved by maxmzng the followng object functon: JðW lda Þ¼ jwt lda S bw lda j jw t lda S ð3þ ww lda j Ths soluton can be shown to correspond to the generalzed egenvectors of the followng equaton: S b w ¼ l S w w ð4þ If S w s a nonsngular matrx then the objecton functon of Eq. (3) s maxmzed when the transformaton matrx W lda conssts of D generalzed egenvectors correspondng to the D largest egenvalues of S 1 w S b [6]. In other words, by sortng the egenvalues n descendng order, we can then use the correspondng frst D egenvectors to form the columns of the LDA matrx W lda. In practcal applcaton, egenvectors wth low egenvalues can be dscarded to compress a hgh dmensonal feature to a lowdmensonal feature wth an enhanced dscrmnaton. Note that there are at most M 1 nonzero generalzed egenvalues, so an upper bound on D s M A general soluton for ncremental LDA The problem of Incremental LDA (ILDA) can be descrbed as follows: when new samples are beng presented, how can we update the correspondng LDA model parameters, ncludng the class mean vector m j, j¼1,2, y, M, mean vector m of all classes, wthn-class scatter matrx S w, and between-class scatter matrx S b. Wth these updated parameters, the new updated LDA transformaton matrx W lda can be computed accordngly. Suppose we have L ncremental samples Y¼{y }(¼1, y, L) n P classes. Wthout loss of generalty, we assume that l of L ncremental samples belong to class C (¼1, y, P). It s worthwhle to note that some of the class C may be newly ntroduced classes. Let m y and my represent the mean vector of class C and all ð2þ S y b ¼ XP ¼ 1 l ðm y my Þðm y my Þ T Snce some classes may or may not be updated by the new samples and some new classes contanng only new samples may be ntroduced, the merged class set X can be therefore dvded nto three parts: updated class set W, no updated class set U, and newly ntroduced class set C. We assume the class number s updated from M to T (TZM, TZP), and the sample number of each class s updated as N 0 ¼ N þl, where ¼1, y, T. It s obvous that, f C AF, l ¼0, and f C AG, N ¼0. It can be easly derved that the updated mean m 0 for class C s updated by m 0 ¼ N m þl m y N 0, ¼ 1,...,T ð9þ and the mean vector m 0 of total samples s m 0 ¼ NmþLmy ð10þ N þl Then the updated between-class scatter matrx S 0 b after ncremental data has been presented as follows: S 0 b ¼ XT ¼ 1 N 0 ðm0 m0 Þðm 0 m0 Þ T The updated wthn-class scatter matrx S 0 w s S 0 w ¼ XT X N0 j ðx 0ðjÞ j ¼ 1 ¼ 1 m 0 j Þðx0ðjÞ m 0 j ÞT ¼ XT S 0 j j ¼ 1 ð8þ ð11þ ð12þ where the tranng samples set s gven by X 0 ¼fx 0 gn þ L ¼ 1 ¼fx g N ¼ 1 þfy g L ¼ 1. It can be derved (see [30] for detals) that S 0 j ¼ XN0 j ðx 0ðjÞ ¼ 1 m 0 j Þðx0ðjÞ m 0 j ÞT ¼ S j þ N jl 2 j ðn j þl j Þ 2 ðmy j m jþðm y j m jþ T N2 j þ Xl ðn j þl j Þ 2 þ l jðl j þ2n j Þ ðn j þl j Þ 2 ¼ 1 Xl ¼ 1 m j Þ m j Þ T m y j ÞðyðjÞ m y j ÞT ð13þ The sum for the last three terms of Eq. (13) s rewrtten as follows (see Proof A n the Appendx): N j l 2 j ðn j þl j Þ 2 ðmy j m jþðm y j m jþ T N2 j þ Xl ðn j þl j Þ 2 ¼ 1 m j Þ m j Þ T

4 L. Jn et al. / Neurocomputng 73 (2010) þ l jðl j þ2n j Þ ðn j þl j Þ 2 Xl m y j ÞðyðjÞ m y j ÞT ¼ 1 ¼ S y j þ N jl j ðn j þl j Þ ðmy j m jþðm y j m jþ T ð14þ From Eqs. (13) and (14), the updated wthn-class matrx (12) can be rewrtten as S 0 w ¼ XT S 0 j j ¼ 1S 0 j ¼ X j AO ¼ X S 0 j þ X S y j þ X N j l j ðn j AO j A O j AO j þl j Þ ðmy j m jþðm y j m jþ T ¼ X j j A ðo GÞS þx S j þ X j A G j A ðo FÞS y j þ X S y j j A F þ X N j l j ðn j AC j þl j Þ ðmy j m jþðm y j m jþ T þ X N j l j ðn j þl j Þ ðmy j m jþðm y j m jþ T j A ðo CÞ It s obvous that P j A ðo CÞ N j l j ðn j þl j Þ ðmy j m jþðm y j m jþ T ¼ 0 P S j ¼ 0 j AG P S y j ¼ 0 j A F ð15þ ð16þ Usng the conclusons of Eq. (16) n (15), the updated wthnclass matrx S 0 w s fnally gven as follows: S 0 w ¼ S w þs y w þ X N j l j ðn j AC j þl j Þ ðmy j m jþðm y j m jþ T ð17þ Once we have updated the between-class scatter matrx S 0 b accordng to Eq. (11) and the wthn-class matrx S 0 w accordng to Eq. (17), we can now get the updated LDA transformaton matrx W 0 lda by conductng the egenvalue decomposton of S0ð 1Þ w S0 b. Compared wth the method proposed n [30], our approach provded a general soluton for the ILDA. In [30], the soluton was separated nto two stuatons dependng on whether the ncremental data were sequence data or chunk data. Furthermore, n ether stuaton, the soluton was separated nto two cases agan accordng to whether the new class s ntroduced or not. If the new class was ntroduced, the number of the new class must be 1. In contrast, our approach can solve all of the above stuatons usng a unform framework wthout restrctng the number of newly ntroduced classes Weghted Incremental LDA for wrter adapton for onlne handwrtten Chnese character recognton For the problem of wrter adapton, the handwrtten character samples of a partcular wrter serve as the ncremental sample set Y ¼fy g L ¼ 1 under the ILDA framework. From Eqs. (9) (12) and (17), we can see that the performance of adaptaton could be affected by the number of new samples used. In general, f the new samples of a partcular wrter for learnng the ILDA model are suffcent, t can be expected that the updated ILDA model would gve an mproved accuracy for the specfc wrter, but may sgnfcantly decrease the accuracy for general wrters. Otherwse, f the updatng samples only make up a small proporton of the total tranng data for updatng the LDA model, the performance mprovement mght not be so sgnfcant n such stuaton though the accuracy loss for the general wrter may be very lttle. However, n practcal applcaton, the amount of data that a partcular wrter provdes s uncertan and varous for dfferent characters. On the other hand, we do not expect to make the adaptaton to a specfc wrter s handwrtng style at the cost of losng too much generalty for other wrter styles or to only have lttle mprovement for a partcular wrter. In other words, a good trade-off between wrter-dependent and wrter-ndependent handwrtng recognton s expected. To acheve the trade-off, we nduce a weghted update mechansm to the ILDA. Suppose we have l of L ncremental samples belongng to class C (¼1,y,P), and the orgnal tranng number of the class C s N, a weghted parameter r s ntroduced to compute the weghted wthn-class scatter matrx S ~ y w of the ncremental samples defned as follows: ~S y w ¼ XP X l Q ðy ðþ j m y ÞðyðÞ j m y ÞT, ¼ 1 j ¼ 1 8 r N >< when N l a0, l a0 where Q ¼ l when N ¼ 0, l a0 >: 0 when l ¼ 0 ð18þ Smlarly, the updated mean vector m 0 of each class and the mean vector m 0 of total samples are modfed accordngly as follows: 8 N m þrn m y ¼ m þrm y >< when N ~m 0 ¼ N ð1þrþ ð1þrþ a0, l a0 ¼ 1,...,T m when N a0, l ¼ 0, >: when N ¼ 0, l a0 m y ~m 0 ¼ NmþðrNÞmy ð1þrþn ¼ mþrmy 1þr ð19þ ð20þ The updated weghted between-class scatter matrx S 0 b s gven as follows: ~S 0 b ¼ XT ¼ 1 Q ðm 0 m0 Þðm 0 m0 Þ T, 8 >< l when N ¼ 0 where Q ¼ N when N ¼ 0 >: ð1þrþn when N a0 ð21þ It s worth notng that the WILDA approach s equal to the ILDA approach under the stuaton l ¼0orN ¼0. Snce no ncremental samples are added to class C when l ¼0, N ¼0 means that the class C s a newly ntroduced class. Accordng to the above equatons, the weghted wthn-class scatter matrx S ~ 0 w can be derved by updatng (18) (20) to Eq. (17) accordngly. Then the weghted ILDA transformaton matrx W ~ 0 lda can be computed by conductng the egenvalue decomposton of S 0ð 1Þ w S0 b. We refer to ths modfed ILDA as WILDA n ths paper. In general, the parameter of r s for purpose of controllng the rato of the partal tranng data, whch are used for updatng LDA parameters to the whole tranng data. In other words, larger weghted parameter r, whch ndcates a larger proporton of wrter-specfc ncremental data for updatng the LDA model parameters, means that the WILDA model turned out to be much more adapted for a partcular wrter s styles. Ths would result n hgher recognton accuracy on the partcular wrter s handwrtten samples, but much more accuracy loss for the general wrter, and vce versa. Therefore, the parameter of r should be carefully chosen to acheve a good trade-off between wrterdependent and wrter-ndependent handwrtng recognton.

5 1618 L. Jn et al. / Neurocomputng 73 (2010) Later n the expermental Secton 4.5, we wll desgn a set of experments to see how ths parameter nfluences the performance of wrter-dependent dataset and wrter-ndependent dataset. 3. Classfer desgn and wrter adaptaton based on ILDA/ WILDA Suppose there are M character classes ffc g M ¼ 1g, each modeled by a prototype,l ¼fm p g, where prototype mp s a D dmensonal vector n some feature space. We use K ¼fm p gm to denote the ¼ 1 set of prototype parameters for the classfer. In ths paper, we use the 8-drectonal feature extracton method proposed by Ba and Huo [1] to extract D 1 raw feature vector x for a gven onlne handwrtten Chnese character sample for whch the orgnal feature dmenson s 512,.e. D 1 ¼512. The D 1 raw features are then transformed nto a new feature vector y of dmenson D n the LDA space by usng a D 1 D LDA transformaton matrx W lda,.e. y ¼ W t lda x, where DrD 1.IfDoD 1, the dmenson reducton s acheved by the LDA. The dagram of our proposed wrter adaptaton for onlne handwrtten Chnese character recognton usng ILDA/WILDA s shown n Fg. 1, whch conssts of the tranng phase to tran a general baselne classfer, the wrter adaptaton phase usng ILDA/WILDA, and the classfcaton phase. In the tranng phase, suppose we have a tranng dataset fx ðjþ g j ¼ 1,2,:::,M of N tranng samples belongng to M classes, ¼ 1,2,:::,N j contrbuted by a large number of wrters. The LDA transformaton matrx W lda s frst learned usng the tranng data, and then the class prototype m p s gven by m p j ¼ 1 N j X N j ¼ 1 W t lda xðjþ ¼ 1 N j W t lda X N j ¼ 1 x ðjþ ¼ W t lda m j ð22þ In the classfcaton phase, the feature vector y n the LDA space s compared wth each of the M character prototypes, and a dscrmnant functon s computed for each class C as follows: g ðx,k,w lda Þ¼ mn:y m p : ¼ mn :W t lda x mp : ð23þ The class that gves the maxmum dscrmnant functon s consdered to be the recognzed class,.e. xac k, f k ¼ argmaxg ðx,k,w lda Þ ð24þ In the wrter adaptaton phase usng the ILDA or WILDA approach, when new handwrtten character samples of a partcular wrter are presented, the LDA transformaton matrx W lda s updated, respectvely, through updatng the between- and wthnclass scatter matrxes accordng to Eqs. (12) and (17) for ILDA, and Eqs. (18) and (21) for WILDA, respectvely. Then the classfer prototype parameter set K ¼fm g M ¼ 1 s updated accordng to Eqs. (19) and (22). 4. Expermental results 4.1. Data preparaton and expermental setup The benchmark dataset used n ths paper comes from the SCUT- COUCH database. It s a revson of SCUT-COUCH2008 [21], whchs now contrbuted by more than 168 partcpants. One characterstc of the SCUT-COUCH dataset s that all the samples were collected n a natural way wthout any gudance or constrant for the wrtng styles, therefore, some of the samples were wrtten cursvely. All characters were wrtten n an unconstraned manner. Ths database s a comprehensve dataset composed of 8 subsets: GB1 (3755 level 1 GB ) smplfed Chnese character, GB2 (3038 level 2 GB ) smplfed Chnese character, tradtonal Chnese character (5041 classes), word (8888 classes), Pnyn (2010 classes), dgt (10 classes), alphabet (52 classes) and symbol (122 classes). The SCUT-COUCH database s avalable at Two subsets of SCUT-COUCH dataset are used n our experments. One s the GB1 subset, whch contans 168 wrters samples of 3755 categores of Chnese characters, and the other s the Word8888 subset, whch conssts of 30 wrters samples of 8888 categores of most frequently used handwrtten words. Fg. 2 shows some typcal handwrtten Chnese character samples from the SCUT-COUCH GB1 subset and Word8888 subset. It s worthwhle to notce that a number of characters appear n dfferent places of dfferent words for many tmes, such as the character appears n the words, etc. Ths ndcates that when the Word8888 data were collected from a partcular wrter, the same character was wrtten for many tmes n accordance wth the context of the word corpus. Table 1 gves the statstcs on the 36 most frequently reused characters to show the frequency of a character repeated n dfferent words. Data collected n ths way provde us wth partcularly realstc wrter-ndependent ncremental handwrtten samples. Tranng samples Feature Extracton LDA Transform Tranng LDA Model { S w, S b, W lda } Incremental samples Adaptaton ILDA/WILDA Incremental learnng of LDA updated Model { S w ', S b ', W lda ' } ILDA: W lda W lda ' WILDA: ~ W lda W' lda Classfer Model {Λ, W lda } Classfcaton Testng sample x Feature Extracton LDA Transform Classfcaton: x C k, f k = arg max g (x, Λ, W lda ) Recognton output Fg. 1. Dagram of the wrter adaptve handwrtng recognton system.

6 L. Jn et al. / Neurocomputng 73 (2010) Fg. 2. Some handwrtten Chnese character samples from the SCUT-COUCH subset GB1 and Word8888: (a) 50 handwrtten samples of three Chnese characters from SCUT-Couch GB1 subset and (b) the handwrtten word samples contrbuted by three dfferent wrters that contan the correspondng three Chnese character, respectvely. Table 1 The statstcs of the top 36 most frequently reused characters.

7 1620 L. Jn et al. / Neurocomputng 73 (2010) All of the handwrtten word samples are manually segmented nto solated characters, whch results n 2078 categores of 19,595 solated Chnese characters, to form a new dataset whch we name t as IncCouchDB. In other words, we have a general dataset that contans 168 sets of 3755 classes of GB1 Chnese character (we refer to t as CouchGB1 thereafter) and a wrterdependent ncremental dataset IncCouchDB that contans 30 sets of 2078 classes wthn GB1 level Chnese characters. The two datasets do not share any common wrters. The dataset CouchGB1 s used to tran/test a baselne general purpose wrter-ndependent LDA classfer, and then the ncremental dataset IncCouchDB s used to tran/test the Incremental LDA model for wrter adaptaton. It should be noted that n our experment, new wrters are ntroduced n the IncCouchDB dataset. To buld a general purpose classfer, we randomly select 134 (or 79.16%) sets of data from the CouchGB1 to buld a wrter ndependent baselne classfer, and then use the remanng 34 (or 20.84%) sets to test the performance of the baselne classfer, as well as to evaluate the nfluence after the adaptaton has been conducted for specfc wrter. For each partcular wrter s handwrtten samples from the IncCouchDB dataset, we randomly select 50% of the data n each category for learnng the Incremental LDA model (ILDA or WILDA), and then use the remanng 50% data to test the wrter adapton performance Baselne performance on CouchGB1 and IncCouchDB before wrter adapton wth dfferent LDA dmenson parameters After the classfer s traned by the 134 sets of CouchGB1 data, ts performance s evaluated on the remanng 34 sets of CouchGB1 data and on the new IncCouchDB dataset (30 sets). Table 2 shows the average recognton rate of the testng sample of CouchGB1, wth dfferent LDA dmensons D. Table 3 shows the recognton accuracy on the dataset of IncCouchDB. Due to the lmted length of the paper, we lst only the accuraces for some typcal wrters, especally the wrters whose samples are hard to be recognzed. From Tables 2 and 3, t can be seen that for CouchGB1 testng dataset, we can acheve as hgh as 94% top 1 recognton accuracy and 99.51% top 10 accuracy when D¼512. However, for the 30 sets of IncCouchDB data, the top 1 and top 10 average recognton rates are only 82.77% and 96.18%, respectvely. We can see that several sets of samples (#5, 13, 16, 18, and 22) are partcularly hard to be recognzed (wth very low top 1 recognton accuraces). Ths may be because of large deformatons and the unconstraned cursve styles of these handwrtten samples. Fg. 3 gves some samples taken from them. From Tables 2 and 3, t can also be seen that the recognton accuraces do not drop down sgnfcantly when the dmenson of LDA space s reduced from 512 to 160 (as less as 0.17% and 0.05% for CouchGB1 and IncCouchDB, respectvely). Therefore, n the followng experments, we set the dmenson D for the reduced LDA space as Performance on IncCouchDB after wrter adapton usng WILDA From Table 3, we can see that the accuracy for the 30 sets of IncCouchDB s not good enough. Ths may be due to the fact that many of the wrtng styles of IncCouchDB are unseen n the tranng dataset. It s expected that through the ncremental LDA learnng on a part of specfc wrter tranng data, the baselne classfer s traned to be adapted to the wrter. In ths experment, we update the LDA model usng the ncremental LDA algorthm on the IncCouchDB data. For each wrter s handwrtten data from the IncCouchDB dataset, one half the handwrtten samples are Table 2 Classfcaton accuracy (%) for the CouchGB1 dataset wth dfferent LDA dmenson D. LDA dmenson Recognton rate (%) Top 1 Top 5 Top 10 Top Fg. 3. Some handwrtten samples taken from #5, 13, 16, 18, and 22 of the IncCouchDB dataset. Table 3 Classfcaton accuracy (%) for the IncCouchDB dataset wth dfferent LDA dmensons D. Wrter # D¼512 D¼256 D¼160 D¼96 Top 1 Top 5 Top 10 Top 1 Top 5 Top 10 Top 1 Top 5 Top 10 Top 1 Top 5 Top y y y y y y y y y y y y y y y y y y y y y y y y y y Average

8 L. Jn et al. / Neurocomputng 73 (2010) Table 4 Performance comparson between wth and wthout adaptatons on the IncCouchDB. Wrter # Before WILDA (%) After WILDA (%) Error rate reducton (%) Top 1 Top 5 Top 10 Top 1 Top 5 Top 10 Top 1 Top 5 Top y y y y y y y y y y y y y y y y y y y y Avarage Table 5 Performance comparson of WILDA aganst ILDA on wrter adaptaton. Adaptaton method Recognton rate (%) Top 1 Top 5 Top 10 Wthout adapton ILDA adaptaton WILDA (r¼0.1) WILDA (r¼0.2) WILDA (r¼0.3) WILDA (r¼0.4) WILDA (r¼0.5) WILDA (r¼0.6) Table 6 Performance comparson of IncCouchDB dataset wth dfferent weghtng parameters r. R Before adaptaton (%) After adaptaton (%) Error rate reducton (%) used as tranng data for learnng the WILDA model, and then the remanng half samples are used to test the performance after the WILDA adaptaton. The performance comparson between wth and wthout adaptatons on the IncCouchDB s gven n Table 4 n whch we set the ncremental weghted parameter r¼0.5 n ths experment. From Table 4, t can be seen that the recognton accuraces for all sets are mproved sgnfcantly. In other words, the error rate s reduced dramatcally. In general, the average top 1 recognton rate s mproved from 82.52% to 92.39%, ndcatng a very hgh error reducton of 56.47%¼((17.48% 7.619%)/17.48%). Takng the fve wrters (#5, 13, 16, 18, and 22) as examples, the recognton rate s mproved from 75.01% to 90.44% after WILDA model adapton for wrter #5, whereas the error rate decreases from 24.99% to 9.56%, resultng n a reducton as large as 61.73% (¼(24.99% 9.56%)/24.99%). Smlarly, t can be seen that the recognton rate s mproved from 69.50% to 82.29% for wrter #13, 67.04% to 84.63% for wrter #16, 59.94% to 78.59% for wrter #18, and 43.43% to 75.47% for wrter #22, respectvely. Overall speakng, the results n Table 4 clearly ndcate that the WILDA learnng algorthm s very useful and effectve for the performance mprovement for wrter-specfc adaptaton Performance comparson of wrter adapton usng WILDA aganst ILDA Ths experment s desgned to examne the performance of wrter adaptaton usng the proposed WILDA algorthm aganst the ILDA algorthm. The expermental results on IncCouchDB dataset are gven n Table 5. From Table 5, t can be seen that by applyng the ILDA wrter adapton learnng, the top 1 recognton rate s mproved from 82.52% to 86.92%, whle by applyng the proposed WILDA wrter adapton learnng algorthm, t s mproved to hgher accuraces rangng from 86.49% to 92.82% accordng to dfferent values of r.it can also be seen that when the weghted parameter r s no less than 0.2, the recognton accuracy of the WILDA wrter adapton approach s sgnfcantly hgher than the ILDA approach. Ths clearly ndcates that the proposed WILDA method s better than the ILDA method for wrter adapton n the applcaton of onlne Chnese handwrtten character recognton Performance comparson on IncCouchDB usng WILDA wth dfferent weghted parameters r Ths experment s desgned to examne the effect of dfferent weghted parameters r on the recognton performance. We set the parameter r from 0.05 to 0.6, and the comparson of average recognton rate of the 30 sets n IncCouchDB dataset before and after adaptaton s shown n Table 6. Clearly, t s observed from Table 6 that our wrter adaptaton method usng WILDA can sgnfcantly reduce the error rate for the 30 sets of wrter-dependent dataset; meanwhle, the recognton accuracy s ncreased n accordance wth the ncrease of the weghted parameter r. Ths s reasonable due to the fact that larger updatng rato r means that the new wrter dependent ncremental data wll have more contrbutons to updatng the WILDA model, resultng n better performance of the correspondng updated classfer on the wrter-specfc data. From Table 6, t can be seen that when the updatng rato s 0.6, the recognton accuracy s mproved from orgnal 82.52% to 92.82% Performance comparson on CouchGB1 usng WILDA and ILDA It seems that the recognton performance after usng the ILDA/ WILDA-based adaptaton always acheves sgnfcant mprovement

9 1622 L. Jn et al. / Neurocomputng 73 (2010) Table 7 Performance comparson of wrter-ndependent CouchGB1 dataset. for all the wrter-specfc datasets n IncCouchDB. However, a queston may be whether such knd of adaptaton has dramatc negatve mpact on the general purpose wrter-ndependent dataset CouchGB1. Ths s partcularly an mportant ssue to be taken for consderaton, because we do not expect the adaptaton to specfc wrter s handwrtng style at the cost of losng much generalty for other wrter styles. To examne how the wrter adapton usng the ILDA/WILDA affects the performance on general wrter-ndependent dataset, the recognton results of ILDA/WILDA based wrter adaptaton approach on the general purpose wrterndependent dataset CouchGB1 after the classfer has been updated usng the specfc user-dependent IncCouchDB have been demonstrated n Table 7. Ths table also shows how the parameter of r nfluences the performance of the WILDA based adaptaton approach on the wrter-ndependent dataset CouchGB1. As shown n Table 7, although the recognton accuraces for the general purpose dataset decrease after the adaptaton, the loss s very small (o3%), especally for ILDA and WILDA wth small values of r. Ths ndcates that the proposed wrter adaptaton methods can sgnfcantly reduce the error rate for wrterdependent dataset. In the mean tme, t has lttle negatve mpact on a wrter-ndependent testng dataset. From Table 7, t can be observed that when ro0.4, the accuracy loss on wrter-ndependent dataset s much smaller (o1%). When rz0.4, the proposed method may lose more than 1% accuracy on wrter-ndependent testng dataset. However, such quantty of loss s acceptable (less than 3% even for large updatng rato r). Comparng Table 6 wth Table 7, t can be found that: (1) the mprovement of recognton accuracy for wrter-specfc dataset s much more sgnfcant than the accuracy loss for general wrter-ndependent dataset; (2) the larger the updatng rato r s, the hgher recognton accuracy s for wrter-specfc data, and the lower but acceptable recognton accuracy s for general wrter-ndependent dataset; and (3) a trade-off should be found between the performance for a specfc user dataset and that of the general dataset. In a practcal vew, we suggest that the reasonable range of r should be taken from 0.1 to 0.3. Under such settngs, the proposed adaptaton method can reduce about % error rate on the wrter-dependent dataset whle t has only less than 0.85% accuracy loss on the wrter-ndependent general dataset. 5. Concluson Wthout adaptaton (%) Wth adaptaton (%) Accuracy loss (%) ILDA adaptaton WILDA (r¼0.05) WILDA (r¼0.1) WILDA (r¼0.2) WILDA (r¼0.3) WILDA (r¼0.4) WILDA (r¼0.5) WILDA (r¼0.6) Wrter adaptaton converts a wrter-ndependent system, whch s traned from the data contrbuted by a large group of wrters, to a wrter-dependent system, whch s turned for a partcular wrter usng a specfc ncremental data. Ths adapton has potental advantage of sgnfcantly ncreasng recognton accuraces for a partcular wrter, so t s very useful for a real world applcaton, such as buldng a personalzed onlne handwrtten character nput method. In ths paper, a general soluton for ncremental lnear dscrmnant analyss (ILDA) s presented, and a weghted ncremental lnear dscrmnant analyss (WILDA) method by consderng the ssue of uncertan number of ncremental data for wrter adaptaton s proposed for onlne handwrtten Chnese character recognton. Based on the ncremental learnng of the LDA model usng ILDA or WILDA, the wrter adaptaton s performed by updatng the LDA transformaton matrx and the classfer prototypes n the feature space. From the expermental results on general purpose wrter-ndependent dataset CouchGB1 and wrter-dependent dataset IncCouchDB, we can draw the followng conclusons: (1) Both ILDA and WILDA are very effectve to mprove the recognton accuracy for dfferent partcular wrters. (2) The WILDA outperforms the ILDA for wrter adaptaton. The proposed WILDA method can reduce as much as 47.88% error rate on the IncCouchDB dataset whle t only has as lttle as 0.85% accuracy loss on the CouchGB1 dataset when the weghted parameter r s set to 0.3, showng the effectveness of the proposed wrter adaptaton approach. (3) One good property of our wrter adapton approach based on WILDA s that t can sgnfcantly ncrease the recognton accuracy for the specfc wrter. In the meantme, t has lttle negatve mpact on the performance of general wrters. Acknowledgments We would lke to thank all anonymous revewers for ther valuable suggestons. Ths work s supported n part by NSFC (Grant no. U , ), GDSFC (no ) and Mcrosoft Research Asa (MSAR) fund (Grant no. FY08-RES-THEME-157). Appendx Proof A. N j l 2 j ðn j þl j Þ 2 ðmy j m jþðm y j m jþ T þ N2 j Xl ðn j þl j Þ 2 þ l jðl j þ2n j Þ Xl ðn j þl j Þ 2 ¼ 1 ( 1 ¼ N 2 ðxl ðn j þl j Þ 2 j ¼ 1 m y j ÞðyðjÞ m y j ÞT y ðjþ y ðjþt þl j N j ðl j þn j Þðm j m T j þmy j myt j þl j ðl j þ2n j Þ Xl ¼ 1 ðn j þl j Þ 2 fn2 j þðl 2 j þ2n jl j Þ Xl ¼ 1 X l ¼ 1 l j m y Þ j myt j m y j ÞðyðjÞ m y j ÞT ¼ 1 þl j N j ðl j þn j Þðm j m T j þmy j myt j ¼ 1 m y j mt j m jm yt j Þ ) m y j ÞðyðjÞ m y j ÞT m y j ÞðyðjÞ m y j ÞT g m y j mt j m jm yt j Þ m j Þ m j Þ T 1 ¼ ðn j þl j Þ fðl 2 j þn j Þ 2 S y j þl jn j ðl j þn j Þðm y j m jþðm y j m jþ T g ¼ S y j þ N jl j ðn j þl j Þ ðmy j m jþðm y j m jþ T & References [1] Z. Ba, Q. Huo, A study on the use of 8-drectonal features for onlne handwrtten Chnese character recognton, n: Internatonal Conference on Document Analyss and Recognton (ICDAR), 2005, pp

10 L. Jn et al. / Neurocomputng 73 (2010) [2] C.M. Bshop, Pattern Recognton and Machne Learnng, Sprnger, New York, [3] S.D. Connel, A.K. Jan, Wrter adaptaton of onlne handwrtng models, IEEE Trans. Pattern Anal. Mach. Intell. 24 (3) (2002) [4] H.P. Decell, S.M. Mayekar, Feature combnatons and the dvergence crteron, Comput. Math. Appl. 3 (1977) [5] K. Dng, G. Deng, L. Jn, An nvestgaton of magnary stroke technque for cursve onlne handwrtng Chnese character recognton, n: Internatonal Conference on Document Analyss and Recognton (ICDAR), 2009, pp [6] R.O. Duda, P.E. Hart, D.G. Stork, Pattern Classfcaton, Wley, New York, [7] H. Fu, H. Chang, Y. Xu, H. Pao, User adaptve handwrtng recognton by selfgrowng probablstc decson-based neural networks, IEEE Trans. Neural Networks 11 (6) (2000) [8] H. Fujsawa, Forty years of research n character and document recognton an ndustral perspectve, Pattern Recognton 41 (2008) [9] I. Guyon, C. Warwck, Jont EC US survey of the state-of-the-art n human language technology, / [10] J. Han, B. Bhanu, Indvdual recognton usng gat energy mage, IEEE Trans. Pattern Anal. Mach. Intell. 28 (2) (2006) [11] T. Haste, R. Tbshran, Dscrmnant analyss by gaussan mxtures, J. R. Statst. Soc. Ser. B: Methodologcal 58 (1996) [12] Q. Huo, T. He, A mnmax classfcaton approach to HMM-based onlne handwrtten Chnese character recognton robust aganst affne dstortons, n: Internatonal Conference on Document Analyss and Recognton (ICDAR), 2007, pp [13] A.K. Jan, S. Prabhakar, L. Hong, A multchannel approach to fngerprnt classfcaton, IEEE Trans. Pattern Anal. Mach. Intell. 21 (4) (1999) [14] B. Jelnek, Revew on heteroscedastc dscrmnant analyss, unpublshed report, Center for Advanced Vehcular Systems, Msssspp State Unversty, [15] X. Jng, D. Zhang, A face and palmprnt recognton approach based on dscrmnant DCT feature extracton, IEEE Trans. Syst. Man Cybern. Pt. B 34 (6) (2004) 2,405 2,415. [16] W. Kenzle, K. Chellaplla, Personalzed handwrtng recognton va based regularzaton, n: Internatonal Conference on Machne learnng (ICML), 2006, pp [17] T.K. Km, S.F. Wong, B. Stenger, J. Kttler, R. Cpolla, Incremental lnear dscrmnant analyss usng suffcent spannng set approxmatons, n: Internatonal Conference on Computer Vson and Patter Recognton (CVPR), 2007, pp [18] W.J. Krzanowsk, P. Jonathan, W.V. McCarthy, M.R. Thomas, Dscrmnant analyss wth sngular covarance matrces: methods and applcatons to spectroscopc data, Appl. Stat. 44 (2005) [19] N. Kumar, A.G. Andreou, Heteroscedastc dscrmnant analyss and reduced rank HMMs for mproved speech recognton, Speech Commun. 26 (1998) [20] J.J. LaVola, R.C. Zeleznk, A practcal approach for wrter-dependent symbol recognton usng a wrter-ndependent symbol recognzer, IEEE Trans. Pattern Anal. Mach. Intell. 29 (11) (2007) [21] Y. L, L. Jn, X. Zhu, T. Long, SCUT-COUCH2008: a comprehensve onlne unconstraned Chnese handwrtng dataset, n: Internatonal Conference on Fronters n Handwrtng Recognton (ICFHR), 2008, pp , / [22] C. Lu, S. Jaeger, M. Nakagawa, Onlne recognton of Chnese characters: the state-of-the-art, IEEE Trans. Pattern Anal. Mach. Intell. 26 (2) (2004) [23] C. Lu, X. Zhou, Onlne Japanese character recognton usng trajectory-based normalzaton and drecton feature extracton, a comprehensve onlne unconstraned Chnese handwrtng dataset, n: Internatonal Workshop on Fronters n Handwrtng Recognton (IWFHR), 2006, pp [24] H. Lu, X. Dng, Handwrtten character recognton usng gradent feature and quadratc classfer wth multple dscrmnaton schemes, n: Internatonal Conference on Document Analyss and Recognton (ICDAR), 2005, pp [25] T. Long, L. Jn, A novel orentaton free method for onlne unconstraned cursve handwrtten Chnese word recognton, n: Internatonal Conference on Pattern Recognton (ICPR), 2008, pp [26] M. Loog, R.P.W. Dun, R. Haeb-Umbach, Multclass lnear dmenson reducton by weghted parwse Fsher crtera, IEEE Trans. Pattern Anal. Mach. Intell. 23 (7) (2001) [27] M. Loog, R.P.W. Dun, Lnear dmensonalty reducton va a heteroscedastc extenson of LDA: the Chernoff crteron, IEEE Trans. Pattern Anal. Mach. Intell. 26 (6) (2004) [28] R. Lotlkar, R. Kothar, Fractonal-step dmensonalty reducton, IEEE Trans. Pattern Anal. Mach. Intell. 22 (6) (2000) [29] J. Lu, K.N. Platanots, A.N. Venetsanopoulos, Face recognton usng LDA based algorthms, IEEE Trans. Neural Networks 14 (1) (2003) [30] S. Pang, S. Ozawa, N. Kasabov, Incremental lnear dscrmnant analyss for classfcaton of data streams, IEEE Trans. Syst. Man Cybern. 35 (5) (2005) [31] P.J. Phllps, H. Moon, S.A. Rzv, P.J. Rauss, The FERET evaluaton methodology for face-recognton algorthms, IEEE Trans. Pattern Anal. Mach. Intell. 22 (10) (2000) [32] IEEE Trans. Pattern Anal. Mach. Intell. 22 (1) (2000) [33] M. Szummer, C.M. Bshop, Dscrmnatve wrter adaptaton, n: Internatonal Workshop on Fronters n Handwrtng Recognton (IWFHR), 2006, pp [34] D. Tao, Dscrmnatve lnear and multlnear subspace methods, Ph.D. Thess, School of Computer Scence and Informaton Systems, Brkbeck College, Unversty of London, [35] D. Tao, X. L, X. Wu, S.J. Maybank, General averaged dvergences analyss, n: IEEE Internatonal Conference on Data Mnng, 2007, pp [36] D. Tao, X. L, X. Wu, S.J. Maybank, Geometrc mean for subspace selecton, IEEE Trans. Pattern Anal. Mach. Intell. 31 (2) (2009) [37] D. Tao, X. L, X. Wu, S.J. Maybank, General tensor dscrmnant analyss and Gabor features for gat recognton, IEEE Trans. Pattern Anal. Mach. Intell. 9 (10) (2007) [38] V. Vuor, Adaptve methods for on-lne recognton of solated handwrtten characters, Helsnk Unversty of Technology, Helsnk, [39] D. Wang, C. Lu, J. Yu, X. Zhou, CASIA-OLHWDB1: a database of onlne handwrtten Chnese characters, n: Internatonal Conference on Document Analyss and Recognton (ICDAR), 2009, pp [40] J. Ye, R. Janardan, L. Q, Two-dmensonal lnear dscrmnant analyss, Neural Inform. Process. Syst. 17 (2005) [41] H. Zhao, P.C. Yuen, Incremental lnear dscrmnant analyss for face recognton, IEEE Trans. Syst. Man Cybern. 38 (2008) Lanwen Jn receved a B.S. degree from the Unversty of Scence and Technology of Chna and a Ph.D. degree from South Chna Unversty of Technology n 1991 and 1996, respectvely. He vsted Motorola Chna Research Center n 2000 and the Unversty of Hong Kong n 2002 and 2006 as a research fellow, respectvely. He s now a professor at the School of Electronc and Informaton Engneerng, South Chna Unversty of Technology. He receved the New Century Excellent Talent Program Award of Chna MOE n He has publshed more than 90 papers n the areas of handwrtten Chnese character recognton, mage processng and pattern recognton. Hs current research nterests nclude character recognton, pattern analyss and recognton, mage processng, machne learnng and ntellgent systems. He s a member of the IEEE and IEEE Computer Socety. He served as Program Commttee member for a number of nternatonal conferences, ncludng IWFHR 06, ICFHR 08, ICDAR 09, ICMLC 07, ICMLC 08, ICMLC 09, ICFHR 10. Ka Dng receved hs B.S. degree n 2006, and s a Ph.D. canddate n Communcaton and Informaton System at South Chna Unversty of Technology, Guangzhou, Chna. Hs research nterests nclude pattern recognton, machne learnng and dgtal mage processng. Zhbn Huang receved hs B.S. degree n 2007, and s pursung hs Master s degree, n Communcaton and Informaton System at South Chna Unversty of Technology, Guangzhou, Chna. Hs research nterests nclude pattern recognton, machne learnng and ntellgent system.

Incremental MQDF Learning for Writer Adaptive Handwriting Recognition 1

Incremental MQDF Learning for Writer Adaptive Handwriting Recognition 1 200 2th Internatonal Conference on Fronters n Handwrtng Recognton Incremental MQDF Learnng for Wrter Adaptve Handwrtng Recognton Ka Dng, Lanwen Jn * School of Electronc and Informaton Engneerng, South

More information

Feature Reduction and Selection

Feature Reduction and Selection Feature Reducton and Selecton Dr. Shuang LIANG School of Software Engneerng TongJ Unversty Fall, 2012 Today s Topcs Introducton Problems of Dmensonalty Feature Reducton Statstc methods Prncpal Components

More information

An Empirical Comparative Study of Online Handwriting Chinese Character Recognition:Simplified v.s.traditional

An Empirical Comparative Study of Online Handwriting Chinese Character Recognition:Simplified v.s.traditional 2013 12th Internatonal Conference on Document Analyss and Recognton An Emprcal Comparatve Study of Onlne Handwrtng Chnese Recognton:Smplfed v.s.tradtonal Yan Gao, Lanwen Jn +, Wexn Yang School of Electronc

More information

Learning the Kernel Parameters in Kernel Minimum Distance Classifier

Learning the Kernel Parameters in Kernel Minimum Distance Classifier Learnng the Kernel Parameters n Kernel Mnmum Dstance Classfer Daoqang Zhang 1,, Songcan Chen and Zh-Hua Zhou 1* 1 Natonal Laboratory for Novel Software Technology Nanjng Unversty, Nanjng 193, Chna Department

More information

Term Weighting Classification System Using the Chi-square Statistic for the Classification Subtask at NTCIR-6 Patent Retrieval Task

Term Weighting Classification System Using the Chi-square Statistic for the Classification Subtask at NTCIR-6 Patent Retrieval Task Proceedngs of NTCIR-6 Workshop Meetng, May 15-18, 2007, Tokyo, Japan Term Weghtng Classfcaton System Usng the Ch-square Statstc for the Classfcaton Subtask at NTCIR-6 Patent Retreval Task Kotaro Hashmoto

More information

Parallelism for Nested Loops with Non-uniform and Flow Dependences

Parallelism for Nested Loops with Non-uniform and Flow Dependences Parallelsm for Nested Loops wth Non-unform and Flow Dependences Sam-Jn Jeong Dept. of Informaton & Communcaton Engneerng, Cheonan Unversty, 5, Anseo-dong, Cheonan, Chungnam, 330-80, Korea. seong@cheonan.ac.kr

More information

Content Based Image Retrieval Using 2-D Discrete Wavelet with Texture Feature with Different Classifiers

Content Based Image Retrieval Using 2-D Discrete Wavelet with Texture Feature with Different Classifiers IOSR Journal of Electroncs and Communcaton Engneerng (IOSR-JECE) e-issn: 78-834,p- ISSN: 78-8735.Volume 9, Issue, Ver. IV (Mar - Apr. 04), PP 0-07 Content Based Image Retreval Usng -D Dscrete Wavelet wth

More information

The Research of Support Vector Machine in Agricultural Data Classification

The Research of Support Vector Machine in Agricultural Data Classification The Research of Support Vector Machne n Agrcultural Data Classfcaton Le Sh, Qguo Duan, Xnmng Ma, Me Weng College of Informaton and Management Scence, HeNan Agrcultural Unversty, Zhengzhou 45000 Chna Zhengzhou

More information

A Binarization Algorithm specialized on Document Images and Photos

A Binarization Algorithm specialized on Document Images and Photos A Bnarzaton Algorthm specalzed on Document mages and Photos Ergna Kavalleratou Dept. of nformaton and Communcaton Systems Engneerng Unversty of the Aegean kavalleratou@aegean.gr Abstract n ths paper, a

More information

Human Face Recognition Using Generalized. Kernel Fisher Discriminant

Human Face Recognition Using Generalized. Kernel Fisher Discriminant Human Face Recognton Usng Generalzed Kernel Fsher Dscrmnant ng-yu Sun,2 De-Shuang Huang Ln Guo. Insttute of Intellgent Machnes, Chnese Academy of Scences, P.O.ox 30, Hefe, Anhu, Chna. 2. Department of

More information

Accumulated-Recognition-Rate Normalization for Combining Multiple On/Off-Line Japanese Character Classifiers Tested on a Large Database

Accumulated-Recognition-Rate Normalization for Combining Multiple On/Off-Line Japanese Character Classifiers Tested on a Large Database 4 th Internatonal Workshop on Multple Classfer Systems (MCS23) Guldford, UK Accumulated-Recognton-Rate Normalzaton for Combnng Multple On/Off-Lne Japanese Character Classfers Tested on a Large Database

More information

Face Recognition University at Buffalo CSE666 Lecture Slides Resources:

Face Recognition University at Buffalo CSE666 Lecture Slides Resources: Face Recognton Unversty at Buffalo CSE666 Lecture Sldes Resources: http://www.face-rec.org/algorthms/ Overvew of face recognton algorthms Correlaton - Pxel based correspondence between two face mages Structural

More information

Collaboratively Regularized Nearest Points for Set Based Recognition

Collaboratively Regularized Nearest Points for Set Based Recognition Academc Center for Computng and Meda Studes, Kyoto Unversty Collaboratvely Regularzed Nearest Ponts for Set Based Recognton Yang Wu, Mchhko Mnoh, Masayuk Mukunok Kyoto Unversty 9/1/013 BMVC 013 @ Brstol,

More information

Support Vector Machines

Support Vector Machines /9/207 MIST.6060 Busness Intellgence and Data Mnng What are Support Vector Machnes? Support Vector Machnes Support Vector Machnes (SVMs) are supervsed learnng technques that analyze data and recognze patterns.

More information

Cluster Analysis of Electrical Behavior

Cluster Analysis of Electrical Behavior Journal of Computer and Communcatons, 205, 3, 88-93 Publshed Onlne May 205 n ScRes. http://www.scrp.org/ournal/cc http://dx.do.org/0.4236/cc.205.350 Cluster Analyss of Electrcal Behavor Ln Lu Ln Lu, School

More information

Classifier Selection Based on Data Complexity Measures *

Classifier Selection Based on Data Complexity Measures * Classfer Selecton Based on Data Complexty Measures * Edth Hernández-Reyes, J.A. Carrasco-Ochoa, and J.Fco. Martínez-Trndad Natonal Insttute for Astrophyscs, Optcs and Electroncs, Lus Enrque Erro No.1 Sta.

More information

MULTISPECTRAL IMAGES CLASSIFICATION BASED ON KLT AND ATR AUTOMATIC TARGET RECOGNITION

MULTISPECTRAL IMAGES CLASSIFICATION BASED ON KLT AND ATR AUTOMATIC TARGET RECOGNITION MULTISPECTRAL IMAGES CLASSIFICATION BASED ON KLT AND ATR AUTOMATIC TARGET RECOGNITION Paulo Quntlano 1 & Antono Santa-Rosa 1 Federal Polce Department, Brasla, Brazl. E-mals: quntlano.pqs@dpf.gov.br and

More information

Edge Detection in Noisy Images Using the Support Vector Machines

Edge Detection in Noisy Images Using the Support Vector Machines Edge Detecton n Nosy Images Usng the Support Vector Machnes Hlaro Gómez-Moreno, Saturnno Maldonado-Bascón, Francsco López-Ferreras Sgnal Theory and Communcatons Department. Unversty of Alcalá Crta. Madrd-Barcelona

More information

Classifying Acoustic Transient Signals Using Artificial Intelligence

Classifying Acoustic Transient Signals Using Artificial Intelligence Classfyng Acoustc Transent Sgnals Usng Artfcal Intellgence Steve Sutton, Unversty of North Carolna At Wlmngton (suttons@charter.net) Greg Huff, Unversty of North Carolna At Wlmngton (jgh7476@uncwl.edu)

More information

An Optimal Algorithm for Prufer Codes *

An Optimal Algorithm for Prufer Codes * J. Software Engneerng & Applcatons, 2009, 2: 111-115 do:10.4236/jsea.2009.22016 Publshed Onlne July 2009 (www.scrp.org/journal/jsea) An Optmal Algorthm for Prufer Codes * Xaodong Wang 1, 2, Le Wang 3,

More information

Lecture 5: Multilayer Perceptrons

Lecture 5: Multilayer Perceptrons Lecture 5: Multlayer Perceptrons Roger Grosse 1 Introducton So far, we ve only talked about lnear models: lnear regresson and lnear bnary classfers. We noted that there are functons that can t be represented

More information

An Approach to Real-Time Recognition of Chinese Handwritten Sentences

An Approach to Real-Time Recognition of Chinese Handwritten Sentences An Approach to Real-Tme Recognton of Chnese Handwrtten Sentences Da-Han Wang, Cheng-Ln Lu Natonal Laboratory of Pattern Recognton, Insttute of Automaton of Chnese Academy of Scences, Bejng 100190, P.R.

More information

Tsinghua University at TAC 2009: Summarizing Multi-documents by Information Distance

Tsinghua University at TAC 2009: Summarizing Multi-documents by Information Distance Tsnghua Unversty at TAC 2009: Summarzng Mult-documents by Informaton Dstance Chong Long, Mnle Huang, Xaoyan Zhu State Key Laboratory of Intellgent Technology and Systems, Tsnghua Natonal Laboratory for

More information

Face Recognition Based on SVM and 2DPCA

Face Recognition Based on SVM and 2DPCA Vol. 4, o. 3, September, 2011 Face Recognton Based on SVM and 2DPCA Tha Hoang Le, Len Bu Faculty of Informaton Technology, HCMC Unversty of Scence Faculty of Informaton Scences and Engneerng, Unversty

More information

FEATURE EXTRACTION. Dr. K.Vijayarekha. Associate Dean School of Electrical and Electronics Engineering SASTRA University, Thanjavur

FEATURE EXTRACTION. Dr. K.Vijayarekha. Associate Dean School of Electrical and Electronics Engineering SASTRA University, Thanjavur FEATURE EXTRACTION Dr. K.Vjayarekha Assocate Dean School of Electrcal and Electroncs Engneerng SASTRA Unversty, Thanjavur613 41 Jont Intatve of IITs and IISc Funded by MHRD Page 1 of 8 Table of Contents

More information

RECOGNIZING GENDER THROUGH FACIAL IMAGE USING SUPPORT VECTOR MACHINE

RECOGNIZING GENDER THROUGH FACIAL IMAGE USING SUPPORT VECTOR MACHINE Journal of Theoretcal and Appled Informaton Technology 30 th June 06. Vol.88. No.3 005-06 JATIT & LLS. All rghts reserved. ISSN: 99-8645 www.jatt.org E-ISSN: 87-395 RECOGNIZING GENDER THROUGH FACIAL IMAGE

More information

Support Vector Machines

Support Vector Machines Support Vector Machnes Decson surface s a hyperplane (lne n 2D) n feature space (smlar to the Perceptron) Arguably, the most mportant recent dscovery n machne learnng In a nutshell: map the data to a predetermned

More information

Problem Definitions and Evaluation Criteria for Computational Expensive Optimization

Problem Definitions and Evaluation Criteria for Computational Expensive Optimization Problem efntons and Evaluaton Crtera for Computatonal Expensve Optmzaton B. Lu 1, Q. Chen and Q. Zhang 3, J. J. Lang 4, P. N. Suganthan, B. Y. Qu 6 1 epartment of Computng, Glyndwr Unversty, UK Faclty

More information

UB at GeoCLEF Department of Geography Abstract

UB at GeoCLEF Department of Geography   Abstract UB at GeoCLEF 2006 Mguel E. Ruz (1), Stuart Shapro (2), June Abbas (1), Slva B. Southwck (1) and Davd Mark (3) State Unversty of New York at Buffalo (1) Department of Lbrary and Informaton Studes (2) Department

More information

Optimizing Document Scoring for Query Retrieval

Optimizing Document Scoring for Query Retrieval Optmzng Document Scorng for Query Retreval Brent Ellwen baellwe@cs.stanford.edu Abstract The goal of ths project was to automate the process of tunng a document query engne. Specfcally, I used machne learnng

More information

Determining the Optimal Bandwidth Based on Multi-criterion Fusion

Determining the Optimal Bandwidth Based on Multi-criterion Fusion Proceedngs of 01 4th Internatonal Conference on Machne Learnng and Computng IPCSIT vol. 5 (01) (01) IACSIT Press, Sngapore Determnng the Optmal Bandwdth Based on Mult-crteron Fuson Ha-L Lang 1+, Xan-Mn

More information

Subspace clustering. Clustering. Fundamental to all clustering techniques is the choice of distance measure between data points;

Subspace clustering. Clustering. Fundamental to all clustering techniques is the choice of distance measure between data points; Subspace clusterng Clusterng Fundamental to all clusterng technques s the choce of dstance measure between data ponts; D q ( ) ( ) 2 x x = x x, j k = 1 k jk Squared Eucldean dstance Assumpton: All features

More information

Outline. Discriminative classifiers for image recognition. Where in the World? A nearest neighbor recognition example 4/14/2011. CS 376 Lecture 22 1

Outline. Discriminative classifiers for image recognition. Where in the World? A nearest neighbor recognition example 4/14/2011. CS 376 Lecture 22 1 4/14/011 Outlne Dscrmnatve classfers for mage recognton Wednesday, Aprl 13 Krsten Grauman UT-Austn Last tme: wndow-based generc obect detecton basc ppelne face detecton wth boostng as case study Today:

More information

Wishing you all a Total Quality New Year!

Wishing you all a Total Quality New Year! Total Qualty Management and Sx Sgma Post Graduate Program 214-15 Sesson 4 Vnay Kumar Kalakband Assstant Professor Operatons & Systems Area 1 Wshng you all a Total Qualty New Year! Hope you acheve Sx sgma

More information

BAYESIAN MULTI-SOURCE DOMAIN ADAPTATION

BAYESIAN MULTI-SOURCE DOMAIN ADAPTATION BAYESIAN MULTI-SOURCE DOMAIN ADAPTATION SHI-LIANG SUN, HONG-LEI SHI Department of Computer Scence and Technology, East Chna Normal Unversty 500 Dongchuan Road, Shangha 200241, P. R. Chna E-MAIL: slsun@cs.ecnu.edu.cn,

More information

The Greedy Method. Outline and Reading. Change Money Problem. Greedy Algorithms. Applications of the Greedy Strategy. The Greedy Method Technique

The Greedy Method. Outline and Reading. Change Money Problem. Greedy Algorithms. Applications of the Greedy Strategy. The Greedy Method Technique //00 :0 AM Outlne and Readng The Greedy Method The Greedy Method Technque (secton.) Fractonal Knapsack Problem (secton..) Task Schedulng (secton..) Mnmum Spannng Trees (secton.) Change Money Problem Greedy

More information

Detection of an Object by using Principal Component Analysis

Detection of an Object by using Principal Component Analysis Detecton of an Object by usng Prncpal Component Analyss 1. G. Nagaven, 2. Dr. T. Sreenvasulu Reddy 1. M.Tech, Department of EEE, SVUCE, Trupath, Inda. 2. Assoc. Professor, Department of ECE, SVUCE, Trupath,

More information

SLAM Summer School 2006 Practical 2: SLAM using Monocular Vision

SLAM Summer School 2006 Practical 2: SLAM using Monocular Vision SLAM Summer School 2006 Practcal 2: SLAM usng Monocular Vson Javer Cvera, Unversty of Zaragoza Andrew J. Davson, Imperal College London J.M.M Montel, Unversty of Zaragoza. josemar@unzar.es, jcvera@unzar.es,

More information

Skew Angle Estimation and Correction of Hand Written, Textual and Large areas of Non-Textual Document Images: A Novel Approach

Skew Angle Estimation and Correction of Hand Written, Textual and Large areas of Non-Textual Document Images: A Novel Approach Angle Estmaton and Correcton of Hand Wrtten, Textual and Large areas of Non-Textual Document Images: A Novel Approach D.R.Ramesh Babu Pyush M Kumat Mahesh D Dhannawat PES Insttute of Technology Research

More information

X- Chart Using ANOM Approach

X- Chart Using ANOM Approach ISSN 1684-8403 Journal of Statstcs Volume 17, 010, pp. 3-3 Abstract X- Chart Usng ANOM Approach Gullapall Chakravarth 1 and Chaluvad Venkateswara Rao Control lmts for ndvdual measurements (X) chart are

More information

Available online at Available online at Advanced in Control Engineering and Information Science

Available online at   Available online at   Advanced in Control Engineering and Information Science Avalable onlne at wwwscencedrectcom Avalable onlne at wwwscencedrectcom Proceda Proceda Engneerng Engneerng 00 (2011) 15000 000 (2011) 1642 1646 Proceda Engneerng wwwelsevercom/locate/proceda Advanced

More information

Improvement of Spatial Resolution Using BlockMatching Based Motion Estimation and Frame. Integration

Improvement of Spatial Resolution Using BlockMatching Based Motion Estimation and Frame. Integration Improvement of Spatal Resoluton Usng BlockMatchng Based Moton Estmaton and Frame Integraton Danya Suga and Takayuk Hamamoto Graduate School of Engneerng, Tokyo Unversty of Scence, 6-3-1, Nuku, Katsuska-ku,

More information

Problem Set 3 Solutions

Problem Set 3 Solutions Introducton to Algorthms October 4, 2002 Massachusetts Insttute of Technology 6046J/18410J Professors Erk Demane and Shaf Goldwasser Handout 14 Problem Set 3 Solutons (Exercses were not to be turned n,

More information

Modular PCA Face Recognition Based on Weighted Average

Modular PCA Face Recognition Based on Weighted Average odern Appled Scence odular PCA Face Recognton Based on Weghted Average Chengmao Han (Correspondng author) Department of athematcs, Lny Normal Unversty Lny 76005, Chna E-mal: hanchengmao@163.com Abstract

More information

Shape Representation Robust to the Sketching Order Using Distance Map and Direction Histogram

Shape Representation Robust to the Sketching Order Using Distance Map and Direction Histogram Shape Representaton Robust to the Sketchng Order Usng Dstance Map and Drecton Hstogram Department of Computer Scence Yonse Unversty Kwon Yun CONTENTS Revew Topc Proposed Method System Overvew Sketch Normalzaton

More information

BOOSTING CLASSIFICATION ACCURACY WITH SAMPLES CHOSEN FROM A VALIDATION SET

BOOSTING CLASSIFICATION ACCURACY WITH SAMPLES CHOSEN FROM A VALIDATION SET 1 BOOSTING CLASSIFICATION ACCURACY WITH SAMPLES CHOSEN FROM A VALIDATION SET TZU-CHENG CHUANG School of Electrcal and Computer Engneerng, Purdue Unversty, West Lafayette, Indana 47907 SAUL B. GELFAND School

More information

Benchmarking of Update Learning Strategies on Digit Classifier Systems

Benchmarking of Update Learning Strategies on Digit Classifier Systems 2012 Internatonal Conference on Fronters n Handwrtng Recognton Benchmarkng of Update Learnng Strateges on Dgt Classfer Systems D. Barbuzz, D. Impedovo, G. Prlo Dpartmento d Informatca Unverstà degl Stud

More information

BioTechnology. An Indian Journal FULL PAPER. Trade Science Inc.

BioTechnology. An Indian Journal FULL PAPER. Trade Science Inc. [Type text] [Type text] [Type text] ISSN : 0974-74 Volume 0 Issue BoTechnology 04 An Indan Journal FULL PAPER BTAIJ 0() 04 [684-689] Revew on Chna s sports ndustry fnancng market based on market -orented

More information

th International Conference on Document Analysis and Recognition

th International Conference on Document Analysis and Recognition 2013 12th Internatonal Conference on Document Analyss and Recognton Onlne Handwrtten Cursve Word Recognton Usng Segmentaton-free n Combnaton wth P2DBM-MQDF Blan Zhu 1, Art Shvram 2, Srrangaraj Setlur 2,

More information

An Application of the Dulmage-Mendelsohn Decomposition to Sparse Null Space Bases of Full Row Rank Matrices

An Application of the Dulmage-Mendelsohn Decomposition to Sparse Null Space Bases of Full Row Rank Matrices Internatonal Mathematcal Forum, Vol 7, 2012, no 52, 2549-2554 An Applcaton of the Dulmage-Mendelsohn Decomposton to Sparse Null Space Bases of Full Row Rank Matrces Mostafa Khorramzadeh Department of Mathematcal

More information

Virtual Machine Migration based on Trust Measurement of Computer Node

Virtual Machine Migration based on Trust Measurement of Computer Node Appled Mechancs and Materals Onlne: 2014-04-04 ISSN: 1662-7482, Vols. 536-537, pp 678-682 do:10.4028/www.scentfc.net/amm.536-537.678 2014 Trans Tech Publcatons, Swtzerland Vrtual Machne Mgraton based on

More information

6.854 Advanced Algorithms Petar Maymounkov Problem Set 11 (November 23, 2005) With: Benjamin Rossman, Oren Weimann, and Pouya Kheradpour

6.854 Advanced Algorithms Petar Maymounkov Problem Set 11 (November 23, 2005) With: Benjamin Rossman, Oren Weimann, and Pouya Kheradpour 6.854 Advanced Algorthms Petar Maymounkov Problem Set 11 (November 23, 2005) Wth: Benjamn Rossman, Oren Wemann, and Pouya Kheradpour Problem 1. We reduce vertex cover to MAX-SAT wth weghts, such that the

More information

Outline. Type of Machine Learning. Examples of Application. Unsupervised Learning

Outline. Type of Machine Learning. Examples of Application. Unsupervised Learning Outlne Artfcal Intellgence and ts applcatons Lecture 8 Unsupervsed Learnng Professor Danel Yeung danyeung@eee.org Dr. Patrck Chan patrckchan@eee.org South Chna Unversty of Technology, Chna Introducton

More information

Sum of Linear and Fractional Multiobjective Programming Problem under Fuzzy Rules Constraints

Sum of Linear and Fractional Multiobjective Programming Problem under Fuzzy Rules Constraints Australan Journal of Basc and Appled Scences, 2(4): 1204-1208, 2008 ISSN 1991-8178 Sum of Lnear and Fractonal Multobjectve Programmng Problem under Fuzzy Rules Constrants 1 2 Sanjay Jan and Kalash Lachhwan

More information

Type-2 Fuzzy Non-uniform Rational B-spline Model with Type-2 Fuzzy Data

Type-2 Fuzzy Non-uniform Rational B-spline Model with Type-2 Fuzzy Data Malaysan Journal of Mathematcal Scences 11(S) Aprl : 35 46 (2017) Specal Issue: The 2nd Internatonal Conference and Workshop on Mathematcal Analyss (ICWOMA 2016) MALAYSIAN JOURNAL OF MATHEMATICAL SCIENCES

More information

An Indian Journal FULL PAPER ABSTRACT KEYWORDS. Trade Science Inc.

An Indian Journal FULL PAPER ABSTRACT KEYWORDS. Trade Science Inc. [Type text] [Type text] [Type text] ISSN : 97-735 Volume Issue 9 BoTechnology An Indan Journal FULL PAPER BTAIJ, (9), [333-3] Matlab mult-dmensonal model-based - 3 Chnese football assocaton super league

More information

Face Recognition Method Based on Within-class Clustering SVM

Face Recognition Method Based on Within-class Clustering SVM Face Recognton Method Based on Wthn-class Clusterng SVM Yan Wu, Xao Yao and Yng Xa Department of Computer Scence and Engneerng Tong Unversty Shangha, Chna Abstract - A face recognton method based on Wthn-class

More information

Meta-heuristics for Multidimensional Knapsack Problems

Meta-heuristics for Multidimensional Knapsack Problems 2012 4th Internatonal Conference on Computer Research and Development IPCSIT vol.39 (2012) (2012) IACSIT Press, Sngapore Meta-heurstcs for Multdmensonal Knapsack Problems Zhbao Man + Computer Scence Department,

More information

User Authentication Based On Behavioral Mouse Dynamics Biometrics

User Authentication Based On Behavioral Mouse Dynamics Biometrics User Authentcaton Based On Behavoral Mouse Dynamcs Bometrcs Chee-Hyung Yoon Danel Donghyun Km Department of Computer Scence Department of Computer Scence Stanford Unversty Stanford Unversty Stanford, CA

More information

A Gradient Difference based Technique for Video Text Detection

A Gradient Difference based Technique for Video Text Detection A Gradent Dfference based Technque for Vdeo Text Detecton Palaahnakote Shvakumara, Trung Quy Phan and Chew Lm Tan School of Computng, Natonal Unversty of Sngapore {shva, phanquyt, tancl }@comp.nus.edu.sg

More information

Mathematics 256 a course in differential equations for engineering students

Mathematics 256 a course in differential equations for engineering students Mathematcs 56 a course n dfferental equatons for engneerng students Chapter 5. More effcent methods of numercal soluton Euler s method s qute neffcent. Because the error s essentally proportonal to the

More information

Outline. Self-Organizing Maps (SOM) US Hebbian Learning, Cntd. The learning rule is Hebbian like:

Outline. Self-Organizing Maps (SOM) US Hebbian Learning, Cntd. The learning rule is Hebbian like: Self-Organzng Maps (SOM) Turgay İBRİKÇİ, PhD. Outlne Introducton Structures of SOM SOM Archtecture Neghborhoods SOM Algorthm Examples Summary 1 2 Unsupervsed Hebban Learnng US Hebban Learnng, Cntd 3 A

More information

A Gradient Difference based Technique for Video Text Detection

A Gradient Difference based Technique for Video Text Detection 2009 10th Internatonal Conference on Document Analyss and Recognton A Gradent Dfference based Technque for Vdeo Text Detecton Palaahnakote Shvakumara, Trung Quy Phan and Chew Lm Tan School of Computng,

More information

Feature Extraction Based on Maximum Nearest Subspace Margin Criterion

Feature Extraction Based on Maximum Nearest Subspace Margin Criterion Neural Process Lett DOI 10.7/s11063-012-9252-y Feature Extracton Based on Maxmum Nearest Subspace Margn Crteron Y Chen Zhenzhen L Zhong Jn Sprnger Scence+Busness Meda New York 2012 Abstract Based on the

More information

Helsinki University Of Technology, Systems Analysis Laboratory Mat Independent research projects in applied mathematics (3 cr)

Helsinki University Of Technology, Systems Analysis Laboratory Mat Independent research projects in applied mathematics (3 cr) Helsnk Unversty Of Technology, Systems Analyss Laboratory Mat-2.08 Independent research projects n appled mathematcs (3 cr) "! #$&% Antt Laukkanen 506 R ajlaukka@cc.hut.f 2 Introducton...3 2 Multattrbute

More information

Competitive Sparse Representation Classification for Face Recognition

Competitive Sparse Representation Classification for Face Recognition Vol. 6, No. 8, 05 Compettve Sparse Representaton Classfcaton for Face Recognton Yng Lu Chongqng Key Laboratory of Computatonal Intellgence Chongqng Unversty of Posts and elecommuncatons Chongqng, Chna

More information

Load Balancing for Hex-Cell Interconnection Network

Load Balancing for Hex-Cell Interconnection Network Int. J. Communcatons, Network and System Scences,,, - Publshed Onlne Aprl n ScRes. http://www.scrp.org/journal/jcns http://dx.do.org/./jcns.. Load Balancng for Hex-Cell Interconnecton Network Saher Manaseer,

More information

The Codesign Challenge

The Codesign Challenge ECE 4530 Codesgn Challenge Fall 2007 Hardware/Software Codesgn The Codesgn Challenge Objectves In the codesgn challenge, your task s to accelerate a gven software reference mplementaton as fast as possble.

More information

3D vector computer graphics

3D vector computer graphics 3D vector computer graphcs Paolo Varagnolo: freelance engneer Padova Aprl 2016 Prvate Practce ----------------------------------- 1. Introducton Vector 3D model representaton n computer graphcs requres

More information

Unsupervised Learning

Unsupervised Learning Pattern Recognton Lecture 8 Outlne Introducton Unsupervsed Learnng Parametrc VS Non-Parametrc Approach Mxture of Denstes Maxmum-Lkelhood Estmates Clusterng Prof. Danel Yeung School of Computer Scence and

More information

12/2/2009. Announcements. Parametric / Non-parametric. Case-Based Reasoning. Nearest-Neighbor on Images. Nearest-Neighbor Classification

12/2/2009. Announcements. Parametric / Non-parametric. Case-Based Reasoning. Nearest-Neighbor on Images. Nearest-Neighbor Classification Introducton to Artfcal Intellgence V22.0472-001 Fall 2009 Lecture 24: Nearest-Neghbors & Support Vector Machnes Rob Fergus Dept of Computer Scence, Courant Insttute, NYU Sldes from Danel Yeung, John DeNero

More information

PRÉSENTATIONS DE PROJETS

PRÉSENTATIONS DE PROJETS PRÉSENTATIONS DE PROJETS Rex Onlne (V. Atanasu) What s Rex? Rex s an onlne browser for collectons of wrtten documents [1]. Asde ths core functon t has however many other applcatons that make t nterestng

More information

Analysis of Continuous Beams in General

Analysis of Continuous Beams in General Analyss of Contnuous Beams n General Contnuous beams consdered here are prsmatc, rgdly connected to each beam segment and supported at varous ponts along the beam. onts are selected at ponts of support,

More information

Signature and Lexicon Pruning Techniques

Signature and Lexicon Pruning Techniques Sgnature and Lexcon Prunng Technques Srnvas Palla, Hansheng Le, Venu Govndaraju Centre for Unfed Bometrcs and Sensors Unversty at Buffalo {spalla2, hle, govnd}@cedar.buffalo.edu Abstract Handwrtten word

More information

Keywords - Wep page classification; bag of words model; topic model; hierarchical classification; Support Vector Machines

Keywords - Wep page classification; bag of words model; topic model; hierarchical classification; Support Vector Machines (IJCSIS) Internatonal Journal of Computer Scence and Informaton Securty, Herarchcal Web Page Classfcaton Based on a Topc Model and Neghborng Pages Integraton Wongkot Srura Phayung Meesad Choochart Haruechayasak

More information

Hermite Splines in Lie Groups as Products of Geodesics

Hermite Splines in Lie Groups as Products of Geodesics Hermte Splnes n Le Groups as Products of Geodescs Ethan Eade Updated May 28, 2017 1 Introducton 1.1 Goal Ths document defnes a curve n the Le group G parametrzed by tme and by structural parameters n the

More information

Design of Structure Optimization with APDL

Design of Structure Optimization with APDL Desgn of Structure Optmzaton wth APDL Yanyun School of Cvl Engneerng and Archtecture, East Chna Jaotong Unversty Nanchang 330013 Chna Abstract In ths paper, the desgn process of structure optmzaton wth

More information

The Comparison of Calibration Method of Binocular Stereo Vision System Ke Zhang a *, Zhao Gao b

The Comparison of Calibration Method of Binocular Stereo Vision System Ke Zhang a *, Zhao Gao b 3rd Internatonal Conference on Materal, Mechancal and Manufacturng Engneerng (IC3ME 2015) The Comparson of Calbraton Method of Bnocular Stereo Vson System Ke Zhang a *, Zhao Gao b College of Engneerng,

More information

Online Detection and Classification of Moving Objects Using Progressively Improving Detectors

Online Detection and Classification of Moving Objects Using Progressively Improving Detectors Onlne Detecton and Classfcaton of Movng Objects Usng Progressvely Improvng Detectors Omar Javed Saad Al Mubarak Shah Computer Vson Lab School of Computer Scence Unversty of Central Florda Orlando, FL 32816

More information

Reducing Frame Rate for Object Tracking

Reducing Frame Rate for Object Tracking Reducng Frame Rate for Object Trackng Pavel Korshunov 1 and We Tsang Oo 2 1 Natonal Unversty of Sngapore, Sngapore 11977, pavelkor@comp.nus.edu.sg 2 Natonal Unversty of Sngapore, Sngapore 11977, oowt@comp.nus.edu.sg

More information

FACE RECOGNITION USING MAP DISCRIMINANT ON YCBCR COLOR SPACE

FACE RECOGNITION USING MAP DISCRIMINANT ON YCBCR COLOR SPACE FAC RCOGNIION USING MAP DISCRIMINAN ON YCBCR COLOR SPAC I Gede Pasek Suta Wjaya lectrcal ngneerng Department, ngneerng Faculty, Mataram Unversty. Jl. Majapaht 62 Mataram, West Nusa enggara, Indonesa. mal:

More information

High-Boost Mesh Filtering for 3-D Shape Enhancement

High-Boost Mesh Filtering for 3-D Shape Enhancement Hgh-Boost Mesh Flterng for 3-D Shape Enhancement Hrokazu Yagou Λ Alexander Belyaev y Damng We z Λ y z ; ; Shape Modelng Laboratory, Unversty of Azu, Azu-Wakamatsu 965-8580 Japan y Computer Graphcs Group,

More information

Recognizing Faces. Outline

Recognizing Faces. Outline Recognzng Faces Drk Colbry Outlne Introducton and Motvaton Defnng a feature vector Prncpal Component Analyss Lnear Dscrmnate Analyss !"" #$""% http://www.nfotech.oulu.f/annual/2004 + &'()*) '+)* 2 ! &

More information

A Unified Framework for Semantics and Feature Based Relevance Feedback in Image Retrieval Systems

A Unified Framework for Semantics and Feature Based Relevance Feedback in Image Retrieval Systems A Unfed Framework for Semantcs and Feature Based Relevance Feedback n Image Retreval Systems Ye Lu *, Chunhu Hu 2, Xngquan Zhu 3*, HongJang Zhang 2, Qang Yang * School of Computng Scence Smon Fraser Unversty

More information

An Image Fusion Approach Based on Segmentation Region

An Image Fusion Approach Based on Segmentation Region Rong Wang, L-Qun Gao, Shu Yang, Yu-Hua Cha, and Yan-Chun Lu An Image Fuson Approach Based On Segmentaton Regon An Image Fuson Approach Based on Segmentaton Regon Rong Wang, L-Qun Gao, Shu Yang 3, Yu-Hua

More information

Some material adapted from Mohamed Younis, UMBC CMSC 611 Spr 2003 course slides Some material adapted from Hennessy & Patterson / 2003 Elsevier

Some material adapted from Mohamed Younis, UMBC CMSC 611 Spr 2003 course slides Some material adapted from Hennessy & Patterson / 2003 Elsevier Some materal adapted from Mohamed Youns, UMBC CMSC 611 Spr 2003 course sldes Some materal adapted from Hennessy & Patterson / 2003 Elsever Scence Performance = 1 Executon tme Speedup = Performance (B)

More information

NAG Fortran Library Chapter Introduction. G10 Smoothing in Statistics

NAG Fortran Library Chapter Introduction. G10 Smoothing in Statistics Introducton G10 NAG Fortran Lbrary Chapter Introducton G10 Smoothng n Statstcs Contents 1 Scope of the Chapter... 2 2 Background to the Problems... 2 2.1 Smoothng Methods... 2 2.2 Smoothng Splnes and Regresson

More information

An Iterative Solution Approach to Process Plant Layout using Mixed Integer Optimisation

An Iterative Solution Approach to Process Plant Layout using Mixed Integer Optimisation 17 th European Symposum on Computer Aded Process Engneerng ESCAPE17 V. Plesu and P.S. Agach (Edtors) 2007 Elsever B.V. All rghts reserved. 1 An Iteratve Soluton Approach to Process Plant Layout usng Mxed

More information

Orthogonal Complement Component Analysis for Positive Samples in SVM Based Relevance Feedback Image Retrieval

Orthogonal Complement Component Analysis for Positive Samples in SVM Based Relevance Feedback Image Retrieval Orthogonal Complement Component Analyss for ostve Samples n SVM Based Relevance Feedback Image Retreval Dacheng Tao and Xaoou Tang Department of Informaton Engneerng The Chnese Unversty of Hong Kong {dctao2,

More information

A Modified Median Filter for the Removal of Impulse Noise Based on the Support Vector Machines

A Modified Median Filter for the Removal of Impulse Noise Based on the Support Vector Machines A Modfed Medan Flter for the Removal of Impulse Nose Based on the Support Vector Machnes H. GOMEZ-MORENO, S. MALDONADO-BASCON, F. LOPEZ-FERRERAS, M. UTRILLA- MANSO AND P. GIL-JIMENEZ Departamento de Teoría

More information

PCA Based Gait Segmentation

PCA Based Gait Segmentation Honggu L, Cupng Sh & Xngguo L PCA Based Gat Segmentaton PCA Based Gat Segmentaton Honggu L, Cupng Sh, and Xngguo L 2 Electronc Department, Physcs College, Yangzhou Unversty, 225002 Yangzhou, Chna 2 Department

More information

Pruning Training Corpus to Speedup Text Classification 1

Pruning Training Corpus to Speedup Text Classification 1 Prunng Tranng Corpus to Speedup Text Classfcaton Jhong Guan and Shugeng Zhou School of Computer Scence, Wuhan Unversty, Wuhan, 430079, Chna hguan@wtusm.edu.cn State Key Lab of Software Engneerng, Wuhan

More information

Backpropagation: In Search of Performance Parameters

Backpropagation: In Search of Performance Parameters Bacpropagaton: In Search of Performance Parameters ANIL KUMAR ENUMULAPALLY, LINGGUO BU, and KHOSROW KAIKHAH, Ph.D. Computer Scence Department Texas State Unversty-San Marcos San Marcos, TX-78666 USA ae049@txstate.edu,

More information

Positive Semi-definite Programming Localization in Wireless Sensor Networks

Positive Semi-definite Programming Localization in Wireless Sensor Networks Postve Sem-defnte Programmng Localzaton n Wreless Sensor etworks Shengdong Xe 1,, Jn Wang, Aqun Hu 1, Yunl Gu, Jang Xu, 1 School of Informaton Scence and Engneerng, Southeast Unversty, 10096, anjng Computer

More information

Learning-Based Top-N Selection Query Evaluation over Relational Databases

Learning-Based Top-N Selection Query Evaluation over Relational Databases Learnng-Based Top-N Selecton Query Evaluaton over Relatonal Databases Lang Zhu *, Wey Meng ** * School of Mathematcs and Computer Scence, Hebe Unversty, Baodng, Hebe 071002, Chna, zhu@mal.hbu.edu.cn **

More information

Learning a Class-Specific Dictionary for Facial Expression Recognition

Learning a Class-Specific Dictionary for Facial Expression Recognition BULGARIAN ACADEMY OF SCIENCES CYBERNETICS AND INFORMATION TECHNOLOGIES Volume 16, No 4 Sofa 016 Prnt ISSN: 1311-970; Onlne ISSN: 1314-4081 DOI: 10.1515/cat-016-0067 Learnng a Class-Specfc Dctonary for

More information

Solving two-person zero-sum game by Matlab

Solving two-person zero-sum game by Matlab Appled Mechancs and Materals Onlne: 2011-02-02 ISSN: 1662-7482, Vols. 50-51, pp 262-265 do:10.4028/www.scentfc.net/amm.50-51.262 2011 Trans Tech Publcatons, Swtzerland Solvng two-person zero-sum game by

More information

A Novel Adaptive Descriptor Algorithm for Ternary Pattern Textures

A Novel Adaptive Descriptor Algorithm for Ternary Pattern Textures A Novel Adaptve Descrptor Algorthm for Ternary Pattern Textures Fahuan Hu 1,2, Guopng Lu 1 *, Zengwen Dong 1 1.School of Mechancal & Electrcal Engneerng, Nanchang Unversty, Nanchang, 330031, Chna; 2. School

More information

Steps for Computing the Dissimilarity, Entropy, Herfindahl-Hirschman and. Accessibility (Gravity with Competition) Indices

Steps for Computing the Dissimilarity, Entropy, Herfindahl-Hirschman and. Accessibility (Gravity with Competition) Indices Steps for Computng the Dssmlarty, Entropy, Herfndahl-Hrschman and Accessblty (Gravty wth Competton) Indces I. Dssmlarty Index Measurement: The followng formula can be used to measure the evenness between

More information

A Fast Visual Tracking Algorithm Based on Circle Pixels Matching

A Fast Visual Tracking Algorithm Based on Circle Pixels Matching A Fast Vsual Trackng Algorthm Based on Crcle Pxels Matchng Zhqang Hou hou_zhq@sohu.com Chongzhao Han czhan@mal.xjtu.edu.cn Ln Zheng Abstract: A fast vsual trackng algorthm based on crcle pxels matchng

More information