OVERSIZED DOCUMENT COPYING SYSTEM

OVERSIZED DOCUMENT COPYING SYSTEM JOHN F. CULLEN, JONATHAN J. HULL Ricoh California Reearch Center, 2882 San Hill Roa, #115, Menlo Park, CA 94025, USA A ytem of algorithm i reente for coying ocument that are larger than a normal coier laten. The objective i to rouce a ingle outut image by comuting the regitration between a number of image fragment that are canne earately but cover the ocument to be coie. A new metho for ajuting the regitration of multile image fragment i rooe in thi aer. Thi technique avoi the roagation of error that can occur in a equential regitration algorithm by aoting an iterative aroach that effectively itribute regitration error over the fragment. A relate metho i alo icue that regiter image rereente in comree form. Exerimental reult are reente that emontrate the ability of the rooe metho to accurately regiter multile image fragment from large ocument. Thee reult are comare with thoe obtaine with a reviou equential technique. 1 Introuction The coying of ocument larger than the ize of a normal coier laten woul enable the economical rerouction of ma, newaer, an engineering rawing. Current olution require the ue of large format canner that are exenive an ifficult to jutify for mall an meium-ize office. It i often neceary for uer to take uch ocument to ecial coy ho for rerouction. A convenient an inexenive ytem for coying overize ocument woul allow uer to image ortion (i.e., fragment) of a ocument an regiter thoe fragment automatically. A uer of uch a ytem woul can overlaing fragment that cover a ocument (ee examle hown in Fig. 1) an an automatic regitration algorithm woul rouce a ingle image rereenting the original ocument. That image woul either be tile acro multile heet of aer or reuce to fit on a ingle heet. Ieally, a uer coul can the image fragment with little if any contraint on how much the fragment overla an a fully automatic regitration roce woul rouce the final outut image. However, thee are ifficult characteritic to accommoate becaue of the large earch ace that woul nee to be coniere. The olution aote by the algorithm rooe in thi aer require that the fragment overla IAPR Workho on Document Analyi Sytem, Malvern, PA, Oct. 14-16, 1996,. 118-142. 118

by a minimum ecifie amount. Alo, the uer mut rovie an x-y tranlation that aroximately align the fragment. Previou relate work ha been in regitration of aerial image an image of 3- D cene where the objective i to etermine the geometric tranformation that align one image of the ame area with another. Thi tranformation i often eigne to correct for tranlation, rotation, cale change an erective itortion [1]. Variou technique have been eigne to accomlih thi, incluing cro-correlation metho [2] an other technique that ue local image roceing [3.]. Thee metho are effective when the two image being aligne iffer from each other by a mall amount. They tyically comute a meaurement of how well two image match at variou X an Y tranlation. The ecific tranlation that maximize thi meaure of goone-of-fit i the tranformation that regiter the two image. A roblem with technique like thi i the comute-intenive cro-correlation meaure that mut be comute between two image over a range of hift in oition. The two-imenional nature of ocument image an the fact that fragment are uually canne in the ame lane have allowe regitration algorithm to be imlifie. One metho for automatic regitration of ocument image fragment ue fiucial mark lace on the original ocument [4]. The two-imenional nature of ocument image an the fact that fragment are uually canne in the ame lane have allowe regitration algorithm to be imlifie. One metho for automatic INPUT PROCESS OUTPUT Coier Platen Poter Sheet of Paer Multile Scan of original Fig.1 Overize ocument coying ytem. IAPR Workho on Document Analyi Sytem, Malvern, PA, Oct. 14-16, 1996,. 118-142. 119

regitration of ocument image fragment ue fiucial mark lace on the original ocument [4]. Uer can fragment that contain the fiucial mark. The regitration algorithm locate the fiucial mark an ue aociate feature to calculate the require x-y tranlation an rotation. Thi technique can rouce accurate reult. However, it i inconvenient for uer to manually lace fiucial marker on the original ocument. A equential regitration algorithm ha been alie to fragment that were canne (left-to-right) by a han canner [5]. The fragment were regitere in that ame orer. The ytem work well when fragment are canne left-to-right (or equivalently, right-to-left) acro a large format ocument. However, it oe not work well for image that contain multile overlaing fragment becaue of error roagation. b c a Fig.2 Examle of error roagation. Incorrect lacement of fragment a an i oberve after equential regitration of fragment a, b, c, an. Error roagation occur in equential regitration when the mall error that occur when two fragment are regitere accumulate uch that when the firt fragment i comare to the lat fragment in a equence, a ignificant error i oberve. Thi IAPR Workho on Document Analyi Sytem, Malvern, PA, Oct. 14-16, 1996,. 118-142. 120

effect i illutrate in Fig. 2 where four fragment canne from a large ma are hown after a equential regitration algorithm wa alie. Fragment a, b, c an were correctly regitere in equence. However, fragment a an coul not be correctly regitere with reect to each other without introucing a ignificant error between fragment a an b. Thi aer rooe an algorithm for regitering multile fragment canne from a large format ocument. An iterative aatation of a equential metho i ue that overcome the roblem of error roagation. Eentially, mall error that occur in the regitration of two fragment are average over all the fragment uch that the error between any two fragment i minimize. Thi alo minimize the ercetibility of uch error. A reliminary invetigation of an efficient metho for regitration of two fragment i alo icue. Thi technique can be alie to run-length encoe comree image. A uch, it i uitable for alication to a limite ubet of the ocument that might be inut to the rooe ytem. The ret of thi aer reent a etaile ecrition of the iterative regitration algorithm. The algorithm i ecribe an examle of it alication are reente. An exerimental analyi of the erformance of the rooe technique i alo given in which the new metho i comare to a reviou equential technique. The ability of the new metho to rouce accetable reult when the reviou metho coul not i hown. The regitration algorithm for comree image i alo ecribe an exerimental reult are reente. Several irection for future reearch are alo given. IAPR Workho on Document Analyi Sytem, Malvern, PA, Oct. 14-16, 1996,. 118-142. 121

2 Prooe Algorithm The rooe iterative algorithm for image regitration i hown in Fig. 3. It wa aume that the fragment were canne with le that 10 egree of kew. After loaing the fragment (te a) an acceting the aroximate lacement of the fragment by the uer (te b), the area of each fragment that overla ome other fragment are etermine an the fragment are orte by the ize of the overla area (te c). The global_error i then calculate a the um of the imilarity ifference (L1-norm, ee equation 2,) within each of the overla (te ). A while loo i then execute (te f-) that imlement the iterative imrovement. Each overla in the riority queue i rocee equentially by the for loo (te h) an a air of interet oint i electe (te i) in each of the fragment that articiate in an overla. In our imlementation the Moravec oerator[7] i ue to chooe the interet oint. Thoe air of interet oint are matche (te j) an the tranlation an rotation neee to align them i calculate (te k). The extraction of interet oint i erforme with the Moravec oerator. (Metho for extracting interet oint are reviewe in Yan[8].) Interet oint are ixel that are within an overla region an have itinctive feature that enable interet oint in one fragment to be matche with the ame oint in overlaing fragment. Thi metho calculate the variance between a central ixel an it north, ALGORITHM ITERATE_MATCH (a) - Loa image fragment (b) - Uer rag an ro fragment in aroximately correct location (c) - Calculate overla an tore in maximum area firt riority queue () - Calculate global_error (e) - Ste_error = 0 (f) - while (te_error < global_error) BEGIN (g) - if (te_error!= 0) global_error = te_error (h) - for each overla, an all image are not regitere BEGIN (i) - elect air of interet oint in overla_1 (j) - earch for match in overla_2 (k) - calculate tranlation an rotation correction (l) - calculate change in te_error (m) - if te_error i reuce (n) - move image_2 to it correct oition (o) - uate te_error an local_error END () - Sort the overla lit with maximum local_error firt END Fig.3 Iterative algorithm for regitration of a et of image fragment. IAPR Workho on Document Analyi Sytem, Malvern, PA, Oct. 14-16, 1996,. 118-142. 122

outh, eat an wet neighbor at a given range (ee equation 1). The interet value of var x y = fxy fx + k y+ l 2 (1) kl S the central ixel i the mean of thee value. Interet oint are extracte from gri cell that are geograhically itribute in the overla area. The interet oint with the highet interet value in each gri cell i ave for further roceing. Any cell with maximum interet value below a Fragment 1 Overla from fragment 1 (enlarge) Overla from fragment 2 (enlarge) Fragment 2 Gri cell an electe interet oint Fig.4 Gri cell from which interet oint are electe. Only the interet oint above a et threhol are retaine. IAPR Workho on Document Analyi Sytem, Malvern, PA, Oct. 14-16, 1996,. 118-142. 123

threhol are neglecte. Of the remaining cell, the one with the maximum value i ave a the rincile interet oint. The econ interet oint in the air i then choen a the one which i furthet away from the rincile oint. Circular temlate urrouning the interet oint an the Eucliean itance between thee two oint are ave an ue later in the matching hae. The temlate iameter i tyically 21 ixel for image canne at 100i, however the iameter nee to be larger for image canne at higher reolution. (ee Fig. 5). Temlate 2 Temlate 1 Angle w.r.t frame of reference Fig.5. Temlate that urroun the electe interet oint. The line egment inicate how the angle of kew of the fragment i calculate The earch algorithm comute the imilarity ifference [2.] of the image in the temlate with the image in the econ fragment. Since the air of fragment may have a rotational mi-alignment, the imilarity ifference ha to be comute with the image rotate at variou angle. If the two oint between the two fragment in the overla are matche correctly, then an etimate of the correction neee to bring the two fragment into alignment by a tranlation an a rotation i comute. A earch roce i then execute that locate the bet match between the air of interet oint in one fragment an the image ata in the other fragment. Since we aume that the uer ha been able to rag an ro the image to within one inch of the correct regitration, the earch i limite to uch an area in the econ fragment of the overla. Each match rooe uring the earch roce i evaluate by a two-te roce. The firt te filter out unlikely matche by alication of a threhol to the outut of he interet oerator. The econ te of evaluating the rooe matche comute the imilarity ifference (L1-norm) between the temlate an the econ fragment (ee equation 2). M N E = 1 -- aij bij ab A (2) i = 1 j = 1 IAPR Workho on Document Analyi Sytem, Malvern, PA, Oct. 14-16, 1996,. 118-142. 124

A a reliability check the Eucliean itance between the interet oint in the firt fragment i comare to the Eucliean itance between the interet oint in the econ fragment. The oint that matche the Eucliean itance between the interet oint are electe a the to caniate oint. If the Eucliean itance oe not match, then the regitration i rejecte. The correction neee to align one fragment with the other i then calculate. The angle of rotation i calculate from the angle between the line that interect each air of oint. One fragment in the overla air i then tranlate an rotate in accorance with the calculate correction. Image fragment are merge after regitration an image ata in the overla i relace with the image ata from one of the fragment. 3 Examle A et of four image fragment canne from an original ocument can be een in Fig 6. Each of thee fragment were canne at 100 i with 256 gray level. The Fig.6 Four fragment canne from an original ocument an loae into the ytem. IAPR Workho on Document Analyi Sytem, Malvern, PA, Oct. 14-16, 1996,. 118-142. 125

fragment have been canne with variou amount of overla. In thi cae all fragment overla each other. The uer through a oint an click evice aroximately align the fragment on a ilay. The reult i hown in Fig. 8. A lit of overlaing fragment i contructe by oberving the location of the fragment on the ilay coorinate ytem. The regitration algorithm then refine each air of fragment alignment in turn. A total alignment error i calculate uring each a through the loo. A the regitration of air of fragment i refine, the total alignment error i reuce. Any change to regitration that increae the total alignment error i rejecte. The overla area are comute between air of the overlaing fragment. Thee overla are then orte bae on their area an are maintaine in a riority queue. One fragment, a member of the overla air articiating in the larget overla i choen a the frame-of-reference for the other fragment. The other fragment in thi air i then regitere with reect to the frame-of-reference fragment. The ytem then move to the next fragment that overla either the frame of reference or a fragment that ha reviouly been regitere. The unregitere fragment of that air i then regitere. Thi rocee until all the fragment have been regitere with reect to each other a well a the frame-of-reference image fragment. Fig.7 Four fragment after manual alignment. IAPR Workho on Document Analyi Sytem, Malvern, PA, Oct. 14-16, 1996,. 118-142. 126

Fig.8 Reult of iterative regitration. The iterative imrovement roce i then alie. Each air-wie overla i viite an an error value i etimate between air of fragment. A total error for all image i alo etimate. The overla with the larget error i now viite an one of the fragment of the air i ajute uch that it reuce the local error of thi air of fragment a well a the global error. The current air of fragment that now contribute the larget amount of error i imilarly correcte. Thi roce continue until a further reuction in global error i no longer oible. In the cae of our examle, the reult can be een in Fig. 8. Thi houl be comare to the reult of a equential regitration alie to the ame ata that i hown in Fig. 2. The ability of the iterative aroach to reuce error roagation i illutrate. IAPR Workho on Document Analyi Sytem, Malvern, PA, Oct. 14-16, 1996,. 118-142. 127

4 Exerimental Reult The image regitration algorithm houl rouce a quantitatively accurate a well a qualitatively leaing reult. The latter i ifficult to evaluate. However the quantitative accuracy of the regitration roce can be meaure againt groun truth ata. In our exeriment the erformance of a equential metho of image regitration wa comare to the rooe iterative technique. Both metho were given the ame rag an ro coorinate. The equential metho mae one a over the lit of fragment. The iterative metho ue a many ae a were neceary to reuce the global error to a minimum. Six ocument were electe for analyzing the erformance. The tye of ocument inclue two marine chart, three newaer age, an a oter. Thee were choen to be rereentative of the kin of ocument normally coie on thi tye of ytem. Four fragment image were canne from each of the ix tet ocument. Each fragment wa canne at 100 i, 8 bit gray cale, on an 8.5x11 flatbe canner an meaure 600 by 700 ixel in ize. The fragment were choen uch that they overlae with each other giving a total of 6 unique overla air within a et of fragment. Two et of correoning oint were manually electe from each fragment air overla. Thee are referre to a the groun truth oint. The ame tarting oint wa given to each algorithm (equential an iterative) by aoting one et of rag an ro coorinate for each fragment. The ifference were then meaure between the groun truth oint in the correoning fragment giving how accurate the fragment were lace by the uer. Thee ifference are taken with reect to a global frame of reference. The angle by which one fragment i out of alignment i calculate by meauring the angle between the interecting line that are rawn through the air of oint. From thi tarting oint the equential regitration algorithm i run. After termination, the ifference with the groun truth oint (tranlation an rotation) are again recore. The fragment were then returne to their rag-an-ro oition an the iterative regitration roce wa tarte. Again, after termination the ifference (tranlation an rotation) between the calculate oition an the groun truth oint were meaure. Each overla ha a tranlation an rotation meaurement that characterize the accuracy by which the fragment have been brought into correonence. Figure 10 an 11 how lot of the tranlation an rotation error between the fragment hown IAPR Workho on Document Analyi Sytem, Malvern, PA, Oct. 14-16, 1996,. 118-142. 128

in Fig. 2 an 8. The reult after rag-an ro (labelle ) how a wie variation in tranlation error in the overla. The equential algorithm (labelle ) uccefully reuce thi error in mot of the overla. An excetion exit at overla normalize error 0.0 0.05 0.10 0.15 0.20 tukuba normalize rotation error w.r.t. groun truth (a) (b) overla (c) Fig.9 Tranlation error. (a) A lot of error between the groun truth oint at each overla when the uer oe rag an ro. (b) A lot of the error after equential regitration. (c) Plot of the error after iterative imrovement algorithm. normalize error 0.0 0.05 0.10 0.15 tukuba normalize rotation error w.r.t. groun truth (a) (b) overla (c) Fig.10 Rotation error. (a) A lot of error between the groun truth oint at each overla when the uer oe rag an ro. (b) A lot of the error after equential regitration. (c) Plot of the error after iterative imrovement algorithm. IAPR Workho on Document Analyi Sytem, Malvern, PA, Oct. 14-16, 1996,. 118-142. 129

3 an 6 where it wa about the ame. It i clearly hown that the iterative technique reuce the total tranlation error acro all the overla. Thi confirm the viual information rovie in Fig. 8 an reinforce the hyothei that error roagation can be reuce by the rooe iterative technique. In the cae of rotation error the reult for rag-an-ro an the equential regitration technique how a wie variation while the iterative metho effectively mooth the error acro the overla. Tranlation error wa lotte veru rotation error (Fig. 11). Thi alo how that the equential algorithm wa not able to atifactorily correct two of the overla. In the equential regitration technique, error were concentrate in the 3r an 6th overla. When iterative regitration terminate, thi error ha been reuce by 7 ixel an 0.5 egree. normalize rotation error 0.0 0.05 0.10 0.15 tukuba normalize error w.r.t. groun truth Uer rag an ro Sequential regitration Iterative imrovement 0.0 0.05 0.10 0.15 0.20 normalize tranlation error Fig.11 Plot of tranlation error veru the rotation error for each of the ix overla. label the error after rag an ro., label error of the equential technique., label error of the iterative metho. Reult for the comlete tet et are hown in the Aenix. The image outut by the equential an iterative technique are hown along the to of each age. The error lot for tranlation an rotation error are hown along the bottom. The reult how that the iterative metho rouce an overall reult that wa at leat a goo a the equential metho in all the tet cae. Proceing time neee to regiter each et of four fragment, (i.e., not incluing the can time an the uer rag an ro) were between 70 an 90 econ on a 50 MIPS Sun Sarctation 10 with 48 Megabyte of main memory. Exact time are not IAPR Workho on Document Analyi Sytem, Malvern, PA, Oct. 14-16, 1996,. 118-142. 130

given ince roceing time for the ame air of image can vary by a much a 10% eening on uer lacement. 5 Regitration of Comree Image Thi ection icue an algorithm that etermine the rotation an vertical tranlation comonent which regiter two binary image. A two-te roce i ue in which the rotation an vertical tranlation are calculate from the horizontal rojection hitogram by rotating the hitogram irectly intea of rotating the image. Thee hitogram are obtaine from the one-imenional Huffman coe verion of the two image. (Feature from imilar fax-comree ata have been ue to etermine ocument kew [6].) A uch, thi technique can be alie irectly to image comree in thi format.the horizontal tranlation i etermine by matching feature oint in the mall earch ace rovie by the firt te of the (a) (b) Fig.12 Original ocument image an it horizontal rojection hitogram (a). The ame image an it rojection hitogram after rotation by 2.30 egree (b). IAPR Workho on Document Analyi Sytem, Malvern, PA, Oct. 14-16, 1996,. 118-142. 131

algorithm. Exerimental reult are reorte for ucceful regitration of image of text an ma image. An algorithm for ocument image regitration i rooe in thi ection that ue horizontal rojection hitogram. Thi technique i bae on the rincile that if two image are minor variation of one another, the rotation an vertical tranlation that regiter the image can be calculate by comaring their horizontal rojection hitogram. An examle of why thi work i hown in Figure 12. A mall ortion of a ocument image i hown in Figure 12 (a). That ame image i hown in Figure 12 (b) after it wa rotate by 2.3 egree an tranlate. The correoning horizontal rojection hitogram are hown below each fragment. It can be een that the horizontal rojection hitogram i a itinctive rereentation for an image. The horizontal rojection hitogram i calculate from a run length encoe rereentation for an image. The articular rereentation ue here i the oneimenional comonent of the Grou 3 facimile coing tanar. The image regitration algorithm (Algorithm PH_MATCH in Fig. 13) i bae ALGORITHM PH_MATCH /* rotate an comare hitogram calculate from RLE image */ - given image fragment 1 an image fragment 2 - comute rojection hitogram for both fragment (i.e., h 1 an h 2 ) - for each rotation angle /* rotate h 1 by */ - for each row r in h 1 - a b/w*tan( ) to each row of h 1 _new in the range [r -w*tan( ), r] - where:w i the with of the original image, b i the number of black ixel in row r, i.e., h 1 [r] /* the origin i in the uer left */ - for each hift = -... + - tore_comat[i].value = comat(hift(h 1 _new, ),h 2 ) - tore_comat[i].theta = - tore_comat[i].hift = ++i - ort(tore_comat, value ) - rint bet match at tore_comat[0].theta egree tore_comat[0].hift vert. tranlation Fig.13 Algorithm PH_MATCH: regitration by rojection hitogram comarion. on the hyothei that a reaonable aroximation to the rojection hitogram for a rotate image can be calculate by rotating the hitogram irectly. The unerlying IAPR Workho on Document Analyi Sytem, Malvern, PA, Oct. 14-16, 1996,. 118-142. 132

aumtion i that the black ixel in each row are uniformly itribute acro the column. The egree to which the rotate hitogram i viually imilar to the hitogram calculate from the rotate image, een on how well thi aumtion i atifie. Algorithm PH_MATCH alo aume that the original image are comree iff. no. y angle it. iff. no. y angle it. 0 24 2.3 125375 5 23 2.3 131482 1 25 2.3 127869 6 23 2.5 133390 2 23 2.4 128106 7 22 2.4 134147 3 24 2.4 129605 8 26 2.2 134267 4 25 2.2 131072 9 26 2.1 135814 (a) (b) (c) Fig.14 (a). Ten bet angle an y tranlation that regiter the image in Figure 12 (a) an (b); (b). rojection hitogram for image in Figure 1 (b); (c). the rojection hitogram for the image in Figure 1 (a) after rotation by 2.3 egree. uing a run length encoe rereentation. However, thi characteritic i only utilize when the rojection hitogram are firt calculate. After that, the rojection hitogram for the firt image fragment h 1 i rotate through a number of angle. Thi i one by evenly itributing the value in each hitogram cell over the hitogram cell in a range aroun the correoning row. The hift an um of quare IAPR Workho on Document Analyi Sytem, Malvern, PA, Oct. 14-16, 1996,. 118-142. 133

ifference roceure fin the arameter of the bet regitration of the two fragment. An examle of the imlementation i hown in Figure 14. The ten bet choice of algorithm PH_MATCH for the rotation an tranlation that regiter the image in Figure 12 (a) with the image in Figure 12 (b) are hown in Figure 14 (a). It i een that the bet match wa foun with a rotation of 2.300 egree an a y tranlation of 24 ixel. Thi rotation i confirme by manual meaurement irectly on the image ata which howe the actual rotation i between 2.27 egree an 2.34 egree a. The rojection hitogram for the image in Figure 12 (b) i hown in Figure 14 (b) an the reult of rotating the hitogram for the original image (hown in Figure 12 (a)) by the bet anwer foun by algorithm PH_MATCH (2.30 egree) i hown in Figure14 (c). Thi how the intuitively aealing reult that the hitogram for the original kewe image i alo viually very imilar to the rotate hitogram. The accuracy of the rooe algorithm for regitration uing horizontal rojection hitogram alignment wa invetigate. It erformance wa etermine by comarion to the erformance of the technique rooe earlier in thi aer. A tet et of ix image fragment were canne from three viually ifferent region of a Jaanee ma at 300 i binary. Thee fragment were automatically regitere by both technique. In all three cae the rotation angle reicte by the rooe technique i within 0.2 an the vertical hift i within one ixel of that choen by the technique ecribe earlier in thi aer. 6 Dicuion an Concluion A ytem for coying of overize (larger than a normal coier laten) ocument wa ecribe. A technique for image regitration wa icue that i the baic technology in thi ytem. The overall ytem i uitable for incluion in a igital hotocoier. An imlementation emontrate that everal tye of large format ocument can be regitere uccefully. The run time reorte for the examle ocument how that the metho can be aate for ue on a walk-u coier. Future work inclue teting with larger et of fragment to etermine when the iterative imrovement metho break own. A referre metho for correcting tranlation an kew woul be to kee a table of all the correction that are neee by each fragment, an when the final regitration ha been etermine, then erform only one rotation an tranlation for each fragment. Thi woul reult in an imrovement in run time an image quality. Further analyi i alo neee concerning the number of iteration that are neee to make ignificant imrovement in the regitration roce. Other metho a Calculate by atan ( y / x) where y wa either 32 or 33 ixel an x wa 808 ixel. IAPR Workho on Document Analyi Sytem, Malvern, PA, Oct. 14-16, 1996,. 118-142. 134

incluing imulate annealing houl be trie an contrate with the reult of iterative imrovement. Acknowlegment Thank are ue to Dr. Peter E. Hart, Dr. Davi Stork, Mr. Mark Peair, Dr. Ahma Zani an Dr. K.V. Praa for their icuion an contribution. Dr. Vikram Chalana, recently grauate from the Univerity of Wahington, wa involve in eveloing the equential algorithm while he worke at Ricoh uring the ummer of 1993. Aenix Each age how the reult of equential an iterative regitration, an lot of the error in each metho. IAPR Workho on Document Analyi Sytem, Malvern, PA, Oct. 14-16, 1996,. 118-142. 135

(a) (b) normalize error 0.05 0.10 0.15 w normalize tranlation error w.r.t. groun truth normalize error 0.0 0.02 0.06 0.10 0.14 w normalize rotation error w.r.t. groun truth overla overla (c) () Fig.15 (a) Image after equential regitration. (b) Image after regitration by iterative imrovement. (c) Grah of tranlation error lotte againt overla number. () Grah of rotation error lotte againt overla number. IAPR Workho on Document Analyi Sytem, Malvern, PA, Oct. 14-16, 1996,. 118-142. 136

(a) (b) normalize error 0.0 0.05 0.10 0.15 0.20 0.25 bay normalize tranlation error w.r.t. groun truth normalize error 0.0 0.05 0.10 0.15 0.20 0.25 bay normalize rotation error w.r.t. groun truth overla overla (c) () Fig.16 (a) Image after equential regitration. (b) Image after regitration by iterative imrovement. (c) Grah of tranlation error lotte againt overla number. () Grah of rotation error lotte againt overla number. IAPR Workho on Document Analyi Sytem, Malvern, PA, Oct. 14-16, 1996,. 118-142. 137

(a) (b) normalize error 0.02 0.04 0.06 0.08 0.10 0.12 0.14 new normalize tranlation error w.r.t. groun truth normalize error 0.0 0.05 0.10 0.15 0.20 new normalize rotation error w.r.t. groun truth overla overla (c) () Fig.17 (a) Image after equential regitration. (b) Image after regitration by iterative imrovement. (c) Grah of tranlation error lotte againt overla number. () Grah of rotation error lotte againt overla number. IAPR Workho on Document Analyi Sytem, Malvern, PA, Oct. 14-16, 1996,. 118-142. 138

(a) (b) normalize error 0.0 0.05 0.10 0.15 0.20 br normalize tranlation error w.r.t. groun truth normalize error 0.0 0.05 0.10 0.15 0.20 0.25 0.30 br normalize rotation error w.r.t. groun truth overla overla (c) () Fig.18 (a) Image after equential regitration. (b) Image after regitration by iterative imrovement. (c) Grah of tranlation error lotte againt overla number. () Grah of rotation error lotte againt overla number. IAPR Workho on Document Analyi Sytem, Malvern, PA, Oct. 14-16, 1996,. 118-142. 139

(a) (b) normalize error (c) 0.0 0.05 0.10 0.15 0.20tukuba normalize tranlation error w.r.t. groun truth overla normalize error 0.0 0.05 0.10 0.15 tukuba normalize rotation error w.r.t. groun truth overla () Fig.19 (a) Image after equential regitration. (b) Image after regitration by iterative imrovement. (c) Grah of tranlation error lotte againt overla number. () Grah of rotation error lotte againt overla number. IAPR Workho on Document Analyi Sytem, Malvern, PA, Oct. 14-16, 1996,. 118-142. 140

(a) (b) normalize error 0.05 0.10 0.15 0.20 text normalize tranlation error w.r.t. groun truth normalize error 0.0 0.05 0.10 0.15 0.20 text normalize rotation error w.r.t. groun truth overla overla (c) () Fig.20 (a) Image after equential regitration. (b) Image after regitration by iterative imrovement. (c) Grah of tranlation error lotte againt overla number. () Grah of rotation error lotte againt overla number. IAPR Workho on Document Analyi Sytem, Malvern, PA, Oct. 14-16, 1996,. 118-142. 141

Reference 1. L.G. Brown, A Survey of Image Regitration Technique, ACM Comuting Survey, Vol. 24, 1992,. 325-376. 2. D.I. Barnea an H.F. Silverman, A Cla of Algorithm for Fat Digital Image Regitration, IEEE Tran. on Comuter, Vol. C-21, 1972,. 179-186. 3. A. Gohtaby, Image Regitration by Local Aroximation Metho, Image Viion Com., Vol. 6, 1988,. 525-533. 4. H. Ie an H. Takea, Image Joining Metho an Sytem, US-Patent 5,140,647, aigne to Hitachi Lt., iue Aug. 18, 1992. 5. D. Pittelou, P. Philli, J. Martin, Aaratu an Metho for Auto-merging Image, International Publication Number WO 93/12501, File by: Logitech Inc., International Publication Date: 24th June 1993. 6. A. Lawrence Sitz, Skew etermination in CCITT grou 4 comree ocument image, Proceeing of the Symoium on Document Analyi an Information Retrieval, La Vega, Nevaa, 1992,. 11-25. 7. H. Moravec, Viual Maing by a robot rover, Proc. 6th Int. Joint Conf. Artificial Intelligence, Vol. 1, Tokyo, Jaan, Augut, 1979. 8. Lu Yan, Interet Oerator an Fat Imlementation, International Archive of Photogrammetry an Remote Sening, Vol. 27-11, Jaan 1988,. 491-500. IAPR Workho on Document Analyi Sytem, Malvern, PA, Oct. 14-16, 1996,. 118-142. 142