Signal & Image Processing : An Inernaional Journal SIPIJ Vol.3 No.5 Ocober 2012 BLOCK-BASED MOTION ESTIMATION USING THE PIXELWISE CLASSIFICATION OF THE MOTION COMPENSATION ERROR Jun-Yong Kim 1 Rae-Hong Park 1 and Seungjoon Yang 2 1 Deparmen o Elecronic Engineering Sogang Universi Seoul Korea {jkimv rhpark}@sogang.ac.kr 2 School o Elecrical and Compuer Engineering Ulsan Naional Insiue o Science and Technolog Ulsan Korea sang@unis.ac.kr ABSTRACT In his paper we propose block-based moion esimaion ME algorihms based on he pielwise classiicaion o wo dieren moion compensaion MC errors: 1 displaced rame dierence DFD and 2 brighness consrain consanc erm BCCT. Block-based ME has drawbacks such as unreliable moion vecors MVs and blocking ariacs especiall in objec boundaries. The proposed block maching algorihm BMA-based mehods aemp o reduce ariacs in objec-boundar blocks caused b incorrec assumpion o a single rigid ranslaional moion. The ield more appropriae MVs in boundar blocks under he assumpion ha here eis up o hree nonoverlapping regions wih dieren moions. The proposed algorihms also reduce he blocking ariac in he convenional BMA in which he overlapped block moion compensaion OBMC is emploed especiall o he seleced regions o preven he degradaion o deails. Eperimenal resuls wih several es sequences show he eeciveness o he proposed algorihms. KEYWORDS Block Maching Algorihm Moion Esimaion Brighness Consanc Consrain Piel Classiicaion Overlapped Block Moion Compensaion 1. INTRODUCTION Moion esimaion ME is one o he well-known mehods or various video processing applicaions. Among a large number o ME approaches block-based ME such as he block maching algorihm BMA [12] has been adoped in a number o inernaional video coding sandards including moion picure epers group MPEG-2/4 and H.26 [3-6]. Block-based ME is racable and simple o implemen wih a lower complei han piel-based ME mehods hus has a large number o applicaions such as inerlaced-o-progressive conversion IPC [7] and rame rae-up conversion FRC [8-10]. Block-based ME reduces he redundanc o he video sequence in he ime domain whereas he discree cosine ransorm DCT reduces he redundanc in he spaial domain. Generall ME b he BMA wih wo successive video rames can be classiied ino wo pes: global and local. The global moion is occurred b camera moions such as ranslaion scale and roaion whereas he local moion is due o moions o individual objecs conained in he video sequence. More han one objec moion can be possible in some blocks and hus he BMA DOI : 10.5121/sipij.2012.3501 1
Signal & Image Processing : An Inernaional Journal SIPIJ Vol.3 No.5 Ocober 2012 usuall has diicul in accurael inding hese local muliple objec moions in video sequences. In block-based ME an image is divided ino a number o blocks o piels wih an assumpion ha each block has a single moion. The opimal moion vecor MV o each block is ound wih he given ME crierion. MV in he block-based ME represens he displacemen o he block in he curren rame wih respec o he corresponding block in he previous rame ha has he smalles maching crierion e.g. mean absolue dierence MAD or mean square error MSE. Though he BMA is simple and hus applicable o various applicaions i has drawbacks such as unreliable MVs and blocking ariacs which degrade visual quali o he processed video [11]. In deecing MVs using he BMA he assumpion ha a block has a single ranslaional moion is no likel o hold especiall in boundar blocks conaining muliple objecs wih dieren moions. To reduce hese problems MPEG-4 visual considers objec-based image processing. A video sequence is considered as a collecion o a single or muliple video objec planes VOPs. A video objec VO ha consrucs a video scene is segmened b shape moion and so on. However i is no eas o accurael erac VOs rom a video sequence. Overlapped block MC OBMC recenl has provided an eecive eension o he convenional block MC BMC [12-20] in which blocks are overlapped wih each oher o reduce he blocking ariacs and residual errors in MC video. The complee esimae o he piel value in he arge block is decided as a linear combinaion o he previous esimae given b he MVs o he arge block and he piel values o neighboring blocks. The noncausal spaial dependenc beween he blocks leads o he ieraive search or he opimal MV. To reduce he esimaion complei modiied nonieraive OBMC schemes [16-17] have been proposed wih he reasonable coding resuls. In his paper we propose ME algorihms ha have he simplici o he BMA b considering up o hree objecs in a block. The consider he moion compensaion MC errors and aemp o reduce blocking ariacs especiall in objec-boundar blocks caused b incorrec assumpion o a single ranslaional objec moion providing beer image quali wih a more appropriae represenaion o MVs. Allowing more han one objec in a block he proposed BMA-based ME algorihms can obain good resuls especiall in boundar blocks wih muliple moions. Also he proposed algorihms use he OBMC or he seleced region o reduce he blocking ariacs wihou he degradaion o deails. The res o he paper is organized as ollows. In Secion 2 we show he block diagram o he proposed BMA-based ME algorihms ollowed b heir deailed descripion. Eperimenal resuls and discussions are shown in Secion 3. Finall conclusions are given in Secion 4. 2. PROPOSED BMA-BASED ME ALGORITHMS In his secion we illusrae he block diagram o he proposed BMA-based ME algorihms and hen describe he algorihmic procedures in deail. The irs sep o he proposed algorihms corresponds o he convenional BMA. Ne our algorihms are urher reined b region-based processing o he MC error. The proposed algorihms reduce he MC error especiall blocking ariacs in boundar blocks caused b unreliable MVs. The also relec he characerisics o a moving objec such as covered and uncovered regions and hus he reined moion vecors are more accurae and consisen o objec s moions. 2
Signal & Image Processing : An Inernaional Journal SIPIJ Vol.3 No.5 Ocober 2012 2.1. Block Diagram o he Proposed Algorihms Figure 1 shows he block diagram o he proposed algorihms. The consis o our seps: blockwise ME/MC pielwise classiicaion region-based ME/MC and MC conidence map. In Figure 1a he irs sep represens he blockwise ME/MC. Blockwise ME means he convenional BMA in which a single ranslaional objec moion is assumed in a block. The MV deeced b blockwise ME ma be incorrec in which he MC error is large especiall in boundar blocks ha conain more han a single objec or moion ielding degraded reconsrucion images. In Figure 1a he image compensaed b he blockwise MV is passed hrough he second sep pielwise classiicaion sep. Generall he inensi dierence beween wo successive rames is small or he saionar background whereas large or moving objecs. The large dierence occurs near boundar piels o a moving objec. The region wih large posiive negaive inensi dierence values corresponds o he covered uncovered par o a moving objec or vice versa. Thus he proposed algorihm pariions a block ino hree pes o non-overlapping regions region wih small rame dierences region wih large posiive rame dierences and region wih large negaive rame dierences based on he aspec o he rame dierence. The accurac o he moion esimaion ME process is reduced i a block consiss o more han wo pes o regions or eample he uncovered region oen has no inormaion in he orward ME. The objecive o he proposed algorihm is o have more reliable ME b considering dierenl each o hese regions in he ME process. The second sep divides each block ino up o hree nonoverlapping regions which will be eplained in deail in Secion 2.2. The pielwise classiicaion is based on he MC error obained in he irs sep. Considering his MC error we can ge beer moion-compensaed images. The oupu o he second sep gives a sequence in which each block conains up o hree dieren regions. The hird sep perorms he region-based ME/MC. In Figure 1b he deails o he hird sep in Figure 1a are shown. Wih up o hree nonoverlapping regions in a block he hird sep uses dieren processes or region sequences R 1 R2 and R 3. For region sequences R 1 and R 2 he regionwise ME/MC which represens he convenional BMA is perormed. Since region sequences R 1 and R 2 almos correspond o he objec-boundar regions anoher ME process separaing he dieren moion regions inds more accurae MVs close o he rue moion. For region sequence R 3 he overlapped MC is perormed o reduce he blocking ariac. 3
Signal & Image Processing : An Inernaional Journal SIPIJ Vol.3 No.5 Ocober 2012 a b Figure 1. Block diagram o he proposed BMA-based ME algorihms: a Overall block diagram and b Deailed block diagram o he region-based ME/MC. Finall he MC conidence map MCCM shows he conidence beween he wo MC processes menioned above. The MC process among he overall process is repeaed wice in he irs and hird seps as illusraed in Figure 1a. We consider he conidence beween hese wo MC processes. Also since he second sep i.e. he pielwise classiicaion aecs he hird sep he MCCM considers he accurac o he pielwise classiicaion. In consequence we can obain a more accurae and naural reconsruced image sequence. Figure 2 shows he absolue MC error o he 5 h rame o he Salesman sequence in which darker piels signi he piels wih large absolue ME errors. The MC error is large in boundaries o objecs as epeced. Also regions having large moions produce large MC errors. 2.2. Descripion o he Proposed Algorihms We propose wo blockwise ME algorihms depending on he epression o he MC error. Le indicae he inensi a piel in he -h rame wih represening he ime emporal ais. The irs proposed algorihm based on he displaced rame dierence DFD error hereaer called he proposed algorihm DFD is described as ollows. In he irs sep o Figure 1 we ind he approimae MV u v o each block b he BMA based on he DFD deined b u v 1 : u v = arg min u v S = arg min u v S { ξ u v } u v 1 1 where ξ represens he sum o absolue dierences SAD as a maching measure and S denoes a se o candidae MVs in he search range. 4
Signal & Image Processing : An Inernaional Journal SIPIJ Vol.3 No.5 Ocober 2012 The BMA assumes ha all he piels in each block have a single rigid ranslaional moion. Thus a block conaining more han one moion produces a large MC error. Especiall in objec boundaries mos o he deeced MVs ield large MC errors which is illusraed in Figure 2. In he second sep o he proposed algorihms in Figure 1a we r o ind more appropriae MVs under he assumpion ha here eis up o hree nonoverlapping regions wih dieren moions in a block. This classiicaion sep uses he DFD beween he original image and he compensaed image. The region classiicaion is done pielwise as in he sliced BMA SBMA [21]: R 1 = R2 = R3 = { d > α} { d < α } { d α} 2 where d = u v 1 denoes he DFD wih he deeced MV v u and α signiies he posiive hreshold. Noe ha he SBMA classiies piels based on he rame dierence FD whereas he proposed algorihm DFD classiies piels based on he DFD. Dominan or secondar local regions near objec boundaries show he DFD larger han α. Region R 3 represens regions wih small inensi changes when moion occurs usuall in he inerior o an objec. Especiall regions R 1 and R 2 correspond o covered and uncovered regions respecivel. Figure 3 shows block classiicaion o 2 in which 8 8 blocks and α = 10 are used. The block wih more han one region is represened b black gra level 0 whereas he block consising o a single region is represened b whie gra level 255. Noe ha he black blocks represen blocks in objec boundaries wih large local moions. Figure 2. Absolue MC error Salesman sequence 5h rame. Figure 3. Block classiicaion Salesman sequence 352 288 5 h rame 8 8 blocks. Figure 4a shows he piels classiied b 2 wih α = 10 in which seleced regions R and 3 1 R2 R are represened b gra levels 0 128 and 255 respecivel. Mos o he black blocks in Figure 3 are classiied as R 1 and R 2 wih up o hree nonoverlapping regions in which each 5
Signal & Image Processing : An Inernaional Journal SIPIJ Vol.3 No.5 Ocober 2012 region has more accurae MVs. Figure 4b shows percenages o each region R 1 R2 and R 3 as a uncion o he rame number o he Salesman sequence wih α = 10. Percenage values o R 1 and R 2 are indicaed along he le verical ais whereas he percenage value o R 3 is indicaed along he righ verical ais. We observe ha mos o he regions 94-99% are included in region R 3 and he regions 1-6% sensiive o moions are included in regions R 1 and R 2. Noe ha regions R 1 and R 2 have a similar percenage in mos o he rames which also can be conirmed in Figure 4a wih he wo regions gra levels 0 and 128 adjacen o each oher. The inluence o he hreshold value on he perormance will be discussed in Secion 2.2. The MC errors are uilized o disinguish he characerisics o an objec and moions giving meaningul classiicaion and hus good resuls. For region sequences R 1 and R 2 more reliable MVs o a block are obained b separael appling he regionwise ME/MC which is equivalen o he convenional BMA o each region. Noe ha he reined ME mehod or each region uses he same block size and search range as he irs ME. However we need o give a dieren process or region sequence R 3. Since he region sequence R 3 nearl corresponds o regions wih small inensi changes when moion occurs anoher ME/MC hardl gives resuls dieren rom hose o he irs sep in Figure 1a. Tha is anoher ME process nearl does no give he reined MVs or his region. Thus or his region we use he OBMC [16-17] wihou a new ME process o reduce he compuaional complei and he blocking ariac. The OBMC o he oher region sequences R 1 and R 2 corresponding o covered and uncovered regions respecivel ma degrade he deails due o he ineracion o he neighboring blocks. Thus he OBMC applied o he speciic region i.e. R 3 region onl reduces he blocking ariac caused b he block-based ME/MC and simulaneousl reduces he degradaion caused b he OBMC [19-20]. Figure 4. Region classiicaion Salesman sequence 352 288: a Region image 5 h rame and b Region raio 50 rames. Finall he MCCM shows he conidence beween he above wo MC processes in he irs and hird seps as illusraed in Figure 1a. The separae ME/MC processes or he each region give more accurae and reliable MVs and piel values. Thus he classiicaion o he region in he second sep aecs he resuls o he region-based ME/MC in he hird sep. The opimal selecion o hreshold α is no eas and he isolaed piel ma be ound or he speciic hreshold value. These isolaed piels ma degrade he resuls o he regionwise ME/MC. To reduce he eec o he hreshold decision and isolaed piels in he region he MCCM is deined b 6
Signal & Image Processing : An Inernaional Journal SIPIJ Vol.3 No.5 Ocober 2012 7 > = ˆ 1 ˆ 1 ˆ ˆ MCCM d d v u d d v u 3 where 1 ˆ ˆ ˆ = v u d denoes he DFD wih he reined MV ˆ ˆ v u obained b he region-based ME in he hird sep. In consequence we obain a more accurae and naural reconsruced image sequence. For a small moion appling he Talor series epansion o he DFD gives 1 v u v u v u v u + + + + = = 4 where and denoe he parial derivaives o wih respec o and respecivel. We can generalize our algorihm under he brighness consanc consrain assumpion BCCT using 4 ielding he proposed algorihm BCCT. In he proposed algorihm BCCT assuming ha he piel inensi is consan along he moion rajecor we eend he assumpion ha inensi o each region in a block is preserved along he moion rajecor. Using he BCCT in 4 we irs deec he approimae MV v u o he block as described in he irs sep o he proposed algorihm DFD. Wih v u he MC errors can be epressed in erms o he BCCT in 4 and each piel in a block is classiied b 2. Figure 5 shows he peak signal o noise raio PSNR o he reconsruced image 5 h rame o he Salesman sequence b our ME mehods as a uncion o. α Noe ha he hree algorihms wo proposed algorihms and SBMA have he similar characerisics. Also he proposed algorihms give a higher PSNR han he SBMA algorihm or mos o he hreshold values. I is possible and advanageous o make he hreshold α adapive o block eaures such as edge inormaion or local variances
Signal & Image Processing : An Inernaional Journal SIPIJ Vol.3 No.5 Ocober 2012 PSNR db 42 41 40 39 PSNR as a uncion o he hreshold Salesman sequence BMA SBMA Proposed DFD Proposed BCCT 38 0 10 20 30 40 50 Threshold value Figure 5. Threshold selecion in he proposed algorihms Salesman sequence 352 288 5 h rame. 3. EXPERIMENTAL RESULTS AND DISCUSSIONS In his secion we show he eeciveness o he proposed algorihms b compuer simulaion wih several es image sequences. Figure 6 shows es image sequences used in eperimens o compare he perormance o he proposed algorihms wih ha o he convenional mehods including he SBMA [21]. Figure 6a shows he 2 nd rame o he 352 240 Fooball sequence consising o 50 rames Figure 6b shows he 30 h rame o he 352 288 Calendar sequence consising o 50 rames and Figure 6c shows he 12 h rame o he 352 288 Salesman sequence consising o 50 rames. Figures 6a and 6b have a lo o local moions whereas Figure 6c conains less local moions. Figure 6b has more deails han Figures 6a and 6c. a b c Figure 6. Image sequences used in eperimens: a Fooball sequence 352 240 2 nd rame b Calendar sequence 352 288 30 h rame and c Salesman sequence 352 288 12 h rame. For perormance comparison o each algorihm Figures 7 8 and 9 show he absolue MC error he reconsrucion image and he enlarged par o he reconsrucion image o Figures 6a 6b and 6c respecivel. Noe ha onl he proposed algorihm DFD is compared in which he proposed algorihm BCCT suiable or small moions gives worse resuls han he proposed algorihm DFD. Figures 7a 7b and 7c show he absolue MC errors b he BMA he 8
Signal & Image Processing : An Inernaional Journal SIPIJ Vol.3 No.5 Ocober 2012 SBMA [21] and he proposed algorihm DFD respecivel in which 8 8 blocks and 31 31 search area are assumed. Figures 7d 7e and 7 illusrae he reconsrucion images o Figure 6a b he BMA he SBMA and he proposed algorihm DFD respecivel. Figures 7g 7h and 7i show he enlarged images o Figures 7d 7e and 7 respecivel. Similarl Figures 8 and 9 are illusraed o compare he perormance o hree algorihms or Figures 6b and 6c respecivel. In Figures 7a 7b and 7c he absolue MC error is illusraed and he darker region represens he larger magniude. Mos o large absolue MC errors in Figures 7a 7b and 7c are ound a boundaries o objecs and absolue MC errors o he proposed algorihm DFD in Figure 7c are he leas o hree absolue MC error images in Figures 7a 7b and 7c. Blocking ariacs near boundaries o objecs are reduced in Figure 7 compared o hose in Figures 7d and 7e. For eas comparison we enlarge a porion which shows large blocking ariacs o he reconsrucion images. Figure 7i b he proposed algorihm DFD shows he leas blocking ariacs among Figures 7g 7h and 7i and especiall he numbers and name on he back and he line o clohes are clearer han hose in Figures 7g and 7h. a b c d e 9
Signal & Image Processing : An Inernaional Journal SIPIJ Vol.3 No.5 Ocober 2012 g h i Figure 7. Fooball sequence 352 240 2 nd rame: a b and c Absolue MC errors b he BMA he SBMA and he proposed algorihm DFD respecivel d e and Images reconsruced b he BMA he SBMA and he proposed algorihm DFD respecivel g h and i Enlarged regions o d e and respecivel. As in Figures 7a 7b and 7c mos o large absolue MC errors in Figures 8a 8b and 8c are ound a boundaries o objecs and absolue MC errors in Figure 8c are smaller han hose in Figures 8a and 8b. In Figures 8g 8h and 8i which are he enlarged images o Figures 8d 8e and 8 respecivel blocking ariacs in Figure 8i are less han hose in Figures 8g and 8h. Since Figure 6b has a number o deails he overall reducion eecs o absolue MC errors in Figures 8g 8h and 8i are less signiican compared wih Figures 7g 7h and 7i. However numbers in he calendar in Figure 8i are cerainl clearer han hose in Figures 8g and 8h. a b c 10
Signal & Image Processing : An Inernaional Journal SIPIJ Vol.3 No.5 Ocober 2012 d e g h i Figure 8. Calendar sequence 352 288 30 h rame: a b and c Absolue MC errors b he BMA he SBMA and he proposed algorihm DFD respecivel d e and Images reconsruced b he BMA he SBMA and he proposed algorihm DFD respecivel g h and i Enlarged regions o d e and respecivel. In Figures 9a 9b and 9c we observe ha he absolue MC errors in Figure 9c are less noiceable han hose in Figures 9a and 9b. Blocking ariacs in Figure 9i are smaller han hose in Figures 9g and 9h. Noe ha he hand o a man a ape and he objec behind he man in Figure 9i are clear. a b c 11
Signal & Image Processing : An Inernaional Journal SIPIJ Vol.3 No.5 Ocober 2012 d e g h i Figure 9. Salesman sequence 352 288 12 h rame: a b and c Absolue MC errors b he BMA he SBMA and he proposed algorihm DFD respecivel d e and Images reconsruced b he BMA he SBMA and he proposed algorihm DFD respecivel g h and i Enlarged regions o d e and respecivel. Figure 10 shows he PSNR comparison o our ME algorihms BMA SBMA and wo proposed algorihms. Figures 10a 10b and 10c are he PSNR graphs or he Fooball Calendar and Salesman sequences respecivel. In all o Figures 10a 10b and 10c he proposed algorihms give higher PSNRs han he convenional algorihms. Especiall he PSNR o he proposed algorihm DFD shows he bes resuls among he algorihms considered or comparison. Noe ha he PSNR o he proposed algorihm DFD is higher han ha o he proposed algorihm BCCT. The proposed algorihm BCCT is suiable or video sequences wih small and simple moions because i is derived using he irs-order Talor series epansion 12
Signal & Image Processing : An Inernaional Journal SIPIJ Vol.3 No.5 Ocober 2012 a v b c Figure 10. Perormance comparison in erms o he PSNR: a Fooball sequence 352 240 50 rames b Calendar sequence 352 288 50 rames and c Salesman sequence 352 288 50 rames. Figure 11 illusraes he PSNR comparison o hree regions R 1 R2 and R 3 b he BMA and he proposed algorihm DFD or he Fooball sequence. The piels belong o regions R 1 and R 2 are regions near he boundaries o he moving objecs. Figures 11a 11b and 11c show he PSNR graphs o regions R 1 R2 and R 3 respecivel. The PSNR dierence beween he BMA and he proposed algorihm DFD is large in Figures 11a and 11b. This ac resuls rom he regionwise ME/MC i.e. he proposed algorihm DFD separael considers he regions ha have 13
Signal & Image Processing : An Inernaional Journal SIPIJ Vol.3 No.5 Ocober 2012 he high and similar moions. And he PSNR dierence beween he BMA and he proposed algorihm DFD is small in Figure 11c compared wih Figures 11a and 11b. Since region R 3 consiss o piels ha have small moions he MVs o hese areas have similar values in boh he BMA and he proposed algorihm DFD. However he OBMC used in his region improves he quali o he reconsrucion o region R 3. The improved quali o each region gives beer quali over he overall reconsrucion image. a b c Figure 11. PSNR comparison o each region Fooball sequence 352 240 50 rames: a Region R 1 b Region R 2 and c Region R 3. Figure 12 shows he PSNR graph in which he perormance enhancemen b he MCCM deined in 3 is illusraed or he Fooball sequence. Eliminaion o isolaed piels gives higher PSNRs. Noe ha MCCM reduces he eecs o he isolaed piels hus improving he accurac o he ME process. 14
Signal & Image Processing : An Inernaional Journal SIPIJ Vol.3 No.5 Ocober 2012 Figure 12. PSNR comparison wih eec o he isolaed piels Fooball sequence 352 240 50 rames. Table 1 compares he compuaion ime o he convenional BMA he SBMA and he proposed algorihm DFD or hree eperimenal images in Figure 6. The relaive compuaion ime lised in Table 1 is deined b he raio o he average compuaion ime o each mehod per rame wih respec o ha o he BMA on he PC wih 3.01GHz Penium IV 1GB RAM Visual C++ compiler. Noe ha he proposed algorihm DFD requires era ime or he pielwise classiicaion sep. Furher research will ocus on he reducion o he compuaional load o he proposed algorihm DFD. Table 1. Comparison o he relaive compuaion ime. BMA SBMA Proposed DFD Remarks Fooball 1.00 1.34 2.21 352 240 50 rames Calendar 1.00 1.36 2.27 352 288 50 rames Salesman 1.00 1.26 2.21 352 288 50 rames Table 2 compares he perormance o he proposed algorihm DFD wih dieren block size and search area or hree es images in Figure 6. Noe ha he perormance is represened in erms o he average PSNR o he sequence o 50 rames each. Table 2. PSNR comparison o he proposed algorihm DFD or dieren block size and search area uni: db. Block size Search area Fooball Calendar Salesman 8 8 17 17 30.421 27.417 40.394 8 8 31 31 31.047 28.090 40.912 16 16 17 17 27.580 25.351 38.875 16 16 31 31 28.162 25.665 39.207 Remarks 352 240 50 rames 352 288 50 rames 352 288 50 rames 15
Signal & Image Processing : An Inernaional Journal SIPIJ Vol.3 No.5 Ocober 2012 4. CONCLUSIONS In his paper we propose wo BMA-based ME algorihms based on pielwise classiicaion o he MC error: 1 DFD and 2 BCCT. We aemp o reduce he blocking ariacs especiall near he boundaries o objecs. The proposed algorihms classi each block ino up o hree nonoverlapping regions b he MC error o he reconsrucion image. The larger he absolue MC error he worse he quali o he resuling image. Thus we classi he piels wih large absolue MC error as signiican elemens o improve he ME/MC perormance and classi hose piels. The classiied regions have more accurae MVs hus we can obain he improved resuls. Also or he piels wih small absolue MC error he OBMC is used o reduce he blocking ariac. The proposed algorihm DFD gives beer resuls han he proposed algorihm BCCT or es sequences wih large moions. Simulaion resuls wih several es sequences show he improved perormance o he proposed algorihms especiall in objec boundaries. Especiall he regions having large MC errors are reconsruced wih relaivel high PSNRs. Also using up o hree nonoverlapping regions he proposed algorihms can eecivel segmen objecs and background. The can be eecivel applied o accurae ME or video-based applicaions. Furher research will be ocused on he eension o he proposed algorihms o color image sequences. ACKNOWLEDGEMENTS This work was suppored in par b Samsung Elecronics Co. Ld. REFERENCES [1] F. H. Jamil A. Chekima R. R. Porle O. Ahmad N. Parimon BMA perormance o video coding or moion esimaion in Proc. 2012 Third In. Con. Inelligen Ssems Modeling and Simulaion ISMS 2012 pp. 287-290. [2] H.M. Musmann P. Pirsch H.J. Gravoer Advances in picure coding Proc. IEEE 73 1985 523 548. [3] ISO/IEC 13818-2 Generic Coding o Moving Picures and Associaed Audio Inormaion 1995. [4] ISO/IEC 14496-2 Inormaion Technolog-Coding o Audio-Visual Objecs Par 2: Visual 1999. [5] ISO/IEC JTC1/SC29/WG11 MPEG93/N457 MPEG-2 Tes Model Version 5 1993. [6] ITU-T Recommendaion H.263 Video Coding or Low Birae Communicaions 1995. [7] S. Yang Y.-Y. Jung Y.H. Lee R.-H. Park Moion compensaion assised moion adapive inerlacedo-progressive conversion IEEE Trans. Circuis Ss. Video Technol. 13 2004 1138 1148. [8] K.A. Bugwadia E.D. Peajan N.N. Puri Progressive-scan rae up-conversion o 24/30 source maerials or HDTV IEEE Trans. Consum. Elecron. 42 1996 312 321. [9] S.-J. Kang S. Yoo and Y. H. Kim Dual moion esimaion or rame rae up-conversion IEEE Trans. Circuis Ss. Video Technol. 20 2010 1909 1914. [10] B.-W. Jeon G.-I. Lee S.-H. Lee R.-H. Park Coarse-o-ine rame inerpolaion or rame raeconversion using pramid srucure IEEE Trans. Consum. Elecron. 49 2003 499 508. [11] S. Fujiwara A. Taguchi Moion-compensaed rame rae up-conversion based on block maching algorihm wih muli-size blocks in Proc. 2005 In. Smp. Inelligen Signal Processing and Communicaion Ssems 2005 pp. 353-356. [12] B. Tao M.T. Orchard A parameric soluion or opimal overlapped block moion compensaion IEEE Trans. Image Proc. 10 2001 341 350. [13] M.T. Orchard G.J. Sullivan Overlapped block moion compensaion: An esimaion-heoreic approach IEEE Trans. Image Proc. 3 1994 693 699. [14] Y.-W. Chen W.-H. Peng Parameric OBMC or piel-adapive emporal predicion on irregular moion sampling grids IEEE Trans. Circuis Ss. Video Technol. 22 2012 113 127. [15] S. Nogaki M. Oha An overlapped block moion compensaion or high quali moion picure coding. in: Proc. IEEE ISCS 1992 pp. 184 187. [16] J.K. Su R.M. Mersereau Moion esimaion mehods or overlapped block moion compensaion IEEE Trans. Image Processing 9 2000 1509-1521. 16
Signal & Image Processing : An Inernaional Journal SIPIJ Vol.3 No.5 Ocober 2012 [17] J.K. Su R.M. Mersereau Non-ieraive rae-consrained moion esimaion or OBMC in: Proc. IEEE ICIP 1997 pp. 33 36. [18] R. Rajagopalan E. Feig M.T. Orchard Moion opimizaion o ordered blocks or overlapped block moion compensaion IEEE Trans. Circuis Ss. Video Technol. 8 1998 119-123. [19] B.-D. Choi J.-W. Han C.-S. Kim S.-J. Ko Moion-compensaed rame inerpolaion using bilaeral moion esimaion and adapive overlapped block moion compensaion IEEE Trans. Circuis Ss. Video Technol. 17 2007 407 416. [20] W. Woo A. Orega Overlapped block dispari compensaion wih adapive windows or sereo image coding IEEE Trans. Circuis Ss. Video Technol. 10 2000 194 200. [21] E.D. Sciascio C. Guarangnella Objec oriened moion esimaion b sliced-block maching algorihms in: Proc. In. Con. Paern Recogniion 2000 pp. 857 860. Auhors Jun-Yong Kim received he B.S. degree rom Sogang Universi in 2004. He is working oward he M.S. degree in elecronic engineering rom Sogang Universi. His curren research ineress are image processing and resoluion enhancemen. Rae-Hong Park was born in Seoul Korea in 1954. He received he B.S. and M.S. degrees in elecronics engineering rom Seoul Naional Universi Seoul Korea in 1976 and 1979 respecivel and he M.S. and Ph.D. degrees in elecrical engineering rom Sanord Universi Sanord CA in 1981 and 1984 respecivel. In 1984 he joined he acul o he Deparmen o Elecronic Engineering School o Engineering Sogang Universi Seoul Korea where he is currenl a Proessor. In 1990 he spen his sabbaical ear as a Visiing Associae Proessor wih he Compuer Vision Laboraor Cener or Auomaion Research Universi o Marland a College Park. In 2001 and 2004 he spen sabbaical semesers a Digial Media Research and Developmen Cener Samsung Elecronics Co. Ld. DTV image/video enhancemen. His curren research ineress are compuer vision paern recogniion and video communicaion. He served as Edior or he Korea Insiue o Telemaics and Elecronics KITE Journal o Elecronics Engineering rom 1995 o 1996. Dr. Park was he recipien o a 1990 Pos-Docoral Fellowship presened b he Korea Science and Engineering Foundaion KOSEF he 1987 Academic Award presened b he KITE and he 2000 Haedong Paper Award presened b he Insiue o Elecronics Engineers o Korea IEEK he 1997 Firs Sogang Academic Award and he 1999 Proessor Achievemen Ecellence Award presened b Sogang Universi. Seungjoon Yang received he B.S. degree rom Seoul Naional Universi Seoul Korea in 1990 and he M.S. and Ph.D. degrees rom he Universi o Wisconsin Madison in 1993 and 2000 respecivel all in elecrical engineering. He was wih he Digial Media Research and Developmen Cener Samsung Elecronics Compan Ld. Seoul rom Sepember 2000 o Augus 2008. He is currenl wih he School o Elecrical and Compuer Engineering Ulsan Naional Insiue o Science and Technolog Ulsan Korea. His curren research ineress include image processing esimaion heor and muli-rae ssems. Dr. Yang received he Samsung Award or he Bes Technolog Achievemen o he Year in 2008 or his work on he premium digial elevision plaorm projec. 17