An Effectve Approach for Vdeo Copy Detecton and Identfcaton of Msbehavng Users S.Sujatha Dhanalakshm Srnvasan College of Engneerng and Technology Mamallapuram, Chenna, Tamlnadu, Inda Abstract Fast development n the feld of multmeda technology has become easer to store and access large amount of vdeo data. Ths technology has edtng and duplcaton of vdeo data that wll cause to volaton of dgtal rghts. So, copy rghts securty becomes a crtcal problem for the huge volume of vdeo data. Ths has led to the requrng Vdeo Copy Detecton has been actvely learned n a large range of the multmeda applcatons. The Auto dual-threshold method s utlzed and t s segmentng the vdeos nto segments wth the content of homogeneous and then the key frame s extract from every segment. The SIFT features are extracted from that segments of key frames. Then Propose an SVD-based technque to match two vdeo frames wth the SIFT pont set descrptors. To obtan the vdeo sequence matchng result propose a graph- based method. It s used to convert the vdeo sequence nto dentfyng the longest path n the frames to dentfy the vdeo matchng-result wth tme constrant. As the Proposed Work, suppose f the query vdeo s matched then server dentfes the unauthorzed user and block that dentfed user at partcular tme. I.e. based on tmestamp, agan the same user queres the same vdeo then that partcular user tmng wll ncreases (.e. 2 n to prevous block tme). Here we addtonally provde opportunty to all users.e. (user s tres to logn 3 attempts). Suppose f attempt of user s crossed the lmt then that partcular user wll dscard from the network. Keywords Vdeo Copy Detecton, SVD-SIFT, Keyframes, Features, Graph Based Sequence method, TmeStamp I. INTRODUCTION The man objectve of the project s to detect whether the query vdeo frames are a copy of a vdeo from the tran vdeo database or not Auto dual Threshold s used to elmnate the redundant frame. SVD-SIFT features are used to compare the two frames features sets ponts. Graph-based vdeo sequence matchng s used to match the query vdeo and tran vdeo. If the query vdeo s matched then server dentfes the unauthorzed user and block that dentfed user at partcular tme. I.e. based on tmestamp. Wth the rapd development of multmeda technologes and meda, the copyrghted materals become easly coped, stored, and dstrbuted over the Internet. Ths stuaton, asde from enablng users to access nformaton easly, causes huge pracy ssues. One possble soluton to dentfy copyrghted meda s watermarkng. Dgtal watermarkng was proposed for copyrght protecton and fngerprntng. The basc dea s to embed nformaton nto the sgnal of the meda (audo, vdeo, or photo). Some watermarks are vsble (e.g., text or logo of the producer or broadcaster), whle others are hdden n the sgnal, whch cannot be perceved by human eye. Today all DVD moves, vdeo games, audo CDs, etc. have fngerprnts that prove the ownershp of the materal. As a dsadvantage, watermarks are generally fragle to vsual transformatons (e.g., reencodng, change of the resoluton/bt rate). For example, hdden data embedded on a move wll probably be lost when the clp s compressed and uploaded to a vdeo sharng web ste. Besdes, temporal nformaton of the vdeo segments (e.g., frame number, tme-code) s also mportant n some applcatons. Watermarkng technque s not desgned to be used for vdeo retreval by queryng wth a sample vdeo clp. Defnton of copy vdeo: A vdeo V1, by means of varous transformatons such as addton, deleton, modfcaton (of aspect, color, contrast, encodng, and so on), camcordng, and so on, s transformed nto another vdeo V2, then vdeo V2 s called a copy of vdeo V1. Content-based copy detecton (CBCD) s ntroduced as an alternatve, or n fact, complementary research felds to watermarkng approach. The man dea of CBCD s that the meda vsually contans enough nformaton for detectng copes. Therefore, the problem of content-based copy detecton s consdered as vdeo smlarty detecton by usng the vsual smlartes of vdeo clps. Server Keyframes are extracted from the reference vdeo database and features are extracted from these keyframes. The extracted features should be robust and effectve to transformatons by whch the vdeo may undergo. Also, the features can be stored n an ndexng structure to make smlarty comparson effcent. Clent Query vdeos are analyzed. Features are extracted from these vdeos and compared to those stored n the reference database. The matchng results are then analyzed and the detecton results are returned. Based on the study, n these transformatons, pcture n pcture s especally dffcult to be detected. And for detectng ths knd of vdeo copes, local feature of SIFT s normally vald. However, matchng based on local features of each frames n two vdeos s n hgh computatonal complexty. In ths paper, we focus on detectng pcture n pcture and propose twn-threshold segmentaton; feature set matchng, and graph-based sequence matchng method. www.jcst.com 863
II.. RELATED WORK An early method based on colour hstogram ntersecton s proposed by Satoh. Yeh and Cheng use a method that parttons the mage nto 4 regons, and extracts a Markov statonary feature (MSF)-extended HSV colour hstogram. Basharat et al. present a vdeo-matchng framework usng spato-temporal segmentaton. A set of features (colour, texture, moton, and SIFT descrptors) s extracted from each segment, and the smlarty between two vdeos s computed wth a bpartte graph and Earth Mover s Dstance (EMD). TABLE 1 LIST OF TRANSFORMATIONS USED IN THE CBVCD TASK # Transformaton detals T1 Camcordng T2 Pcture-n-pcture Type 1 T3 Inserton of patterns (15 dfferent patterns) T4 Strong re-encodng (change of resoluton, btrate) T5 Change of gamma Combnaton of 3 transformatons amongst: T6 blur, gamma, frame droppng, contrast, compresson, rato, nose (A) T7 Combnaton of 5 transformatons amongst (A) Combnaton of 3 transformatons amongst: T8 crop, shft, contrast, capton, flp, nserton of pattern, pcture-n-pcture Type 2 (orgnal vdeo s behnd) (B) T9 Combnaton of 5 transformatons amongst (B) T10 Combnaton of 5 transformatons amongst all the transformatons from 1 to 9 Wu et al. propose that specfc types of vsual features (.e., texture, ntensty, moton, gradent, frequency, nterest pont) should be used for dfferent types of transformatons by a vdeo near-duplcate vdeo matchng system. The methods based on ponts of nterest and ther trajectores are popular n ths feld. Joly et al. present a technque for content-based vdeo dentfcaton based on local fngerprnts. Local fngerprnts are extracted around nterest ponts detected wth Harrs detector, and matched wth an approxmate nearest neghbors search. In the same authors focus on the retreval process of the proposed CBCD scheme by proposng statstcal smlarty search (S3) as a new approxmate search paradgm. In, Joly et al. present dstorton-based probablstc approxmate smlarty search technque (DPS2) to speed-up conventonal technques lke range queres and sequental scan method n a contentbased copy retreval framework. Zhao et al. extract PCA- SIFT descrptors for matchng wth approxmate nearest neghbour search, and tran SVMs to learn matchng patterns. Law-To et al. present a vdeo ndexng approach usng the trajectores of ponts of nterest along the vdeo sequence. They compute temporal contextual nformaton from local descrptors of nterest ponts, and use ths nformaton n a votng functon for matchng vdeo segments. Ren et al. employ a smlar technque by takng nto account spatal and temporal changes of vsual words constructed by SIFT descrptors and bag-of-words approach. Wllams et al. propose a vdeo copy detecton method based on effcently matchng local spatotemporal feature ponts wth a dsk-based ndexng scheme. In general, extractng and matchng ponts of nterest are costly operatons n terms of computaton tme. There are also promsng copy detecton technques based on the smlarty of temporal actvtes of vdeo clps. Mohan presents a vdeo sequence matchng technque that parttons each frame nto 3 x 3 mages and computes ts ordnal measure to form a fngerprnt. The sequences of fngerprnts are compared for vdeo smlarty matchng. Km and Vasudev use ordnal measures of 2 x 2 parttoned mage and consder the results of varous dsplay format conversons, e.g., letter-box, pllar-box. Some vdeo smlarty detecton methods take the advantage of vsual features that can be drectly extracted from compressed vdeos. Ardzzone et al. use MPEG moton vectors as an alternatve to optcal flows, and show that the moton-based vdeo ndexng method they propose does not requre a full. There are numerous descrptors for nearduplcate mage or vdeo detecton avalable n the lterature. Global statstcs, such as color hstograms, are wdely used to effcently work wth a large corpus. These global descrptors are, n general, effcent to compute, compact n storage, but nsuffcently accurate n terms of ther retreval qualty. Alternatvely, local statstcs, such as nterest ponts calculated wth local descrptors, were proposed n. Ths descrpton type s relatvely nvarant and, thus, robust to mage transformatons such as occlusons and croppng. However, local descrptors requre more storage space and matchng between them s computatonally more complex. In the vdeo doman, both global and local descrptors have been extended to ncorporate temporal nformaton. Law-To et al. presented a comparatve study for vdeo copy detecton and concluded that, for small transformatons, temporal ordnal measurements are effectve, whle methods based on local features demonstrate more promsng results n terms of robustness. However, Thomee et al. conducted a large-scale evaluaton of mage copy detecton systems and reached a somewhat dfferent concluson. Ther chosen method that used nterest ponts performed poorly due to ts nablty to fnd smlar sets of ponts between copes. They concluded that ether a smple medan method or the retna method performs the best. To desgn a practcal copy detecton system whch meets the scalablty requrements, a compact, frame-level descrptor that retans the most relevant nformaton, nstead of just sets of nterest pont descrptors, s desrable. Furthermore, frame level descrptors are readly ntegrated nto fast detecton frameworks such as the one presented n. decomposton of the vdeo, and thus, t s computatonally effcent. Bertn et al. present a clp-matchng algorthm that use vdeo fngerprnt based on standard MPEG-7 descrptors. An effectve combnaton of color layout descrptor (CLD), scalable color descrptor (SCD), and edge hstogram descrptor (EHD) forms the fngerprnt. Fngerprnts are extracted from each clp, and they are www.jcst.com 864
compared usng an edt dstance. Sarkar et al. use CLD as vdeo fngerprnts and propose a non-metrc dstance measure to effcently search for matchng vdeos n hghdmensonal space. Hampapur and Bolle made a comparatve analyss of color hstogram-based and edge-based methods for detectng vdeo copes. Another study by Hampapur et al. compares moton drecton, ordnal ntensty sgnature, and color hstogram sgnature matchng technques. As a result of ths study, they conclude that the technques usng ordnal features outperform the others. State-of-the-art copy detecton technques are evaluated n the comparatve study by Law-To et al. Compared descrptors are categorzed nto 2 groups: global and local. Global descrptors use technques based on the temporal actvty, spatal dstrbuton and spato-temporal dstrbuton. Local descrptors compared n ther study are based on extractng Harrs nterest ponts for keyframes wth hgh global ntensty of moton (AJ), for every frame (VCopT), and nterest ponts where mage values have sgnfcant local varatons n both space and tme. It s stated that no sngle technque s optmal for all applcatons; but ordnal temporal measure s very effcent for small. For the dentfcaton of content lnks between vdeo sequences, supportng the range of applcatons descrbed above, Content-Based Copy Detecton (CBCD) s a very relevant tool. Actually, most of the recent vdeo mnng developments just mentoned are CBCD-related methods. By copes we understand potentally transformed versons of orgnal vdeo sequences. The transformatons belong to a large famly and ther ampltude vares sgnfcantly (e.g. Fg. 2). Fg. 2. Copy(left) and Orgnal Content(Rght) III. AUTO DUAL-THRESHOLD METHOD An auto dual-threshold method to elmnate redundant vdeo frames. Ths method cuts contnuous vdeo frames nto vdeo segments by elmnatng temporal redundancy of the vsual nformaton of contnuous vdeo frames. Ths method has the followng two characterstcs. Frst, two thresholds are used. Specf-cally, one threshold s used for detectng abrupt changes of vsual nformaton of frames and another for gradual changes. Second, the values of two thresholds are deter-mned adaptvely accordng to vdeo content. The auto dual-threshold method to elmnate the redundant frames s shown n Fg. 3. Query Vdeo Preproces sng Orgnal Vdeo Accepted Feature Extracton If Matched Block user (Tmestamp) Feature Database Fg. 1. A framework of vdeo copy detecton system. But CBCD methods that are robust to a wde range of transformatons are also computatonally expensve, and the cost of Vdeo Mnng by content-based Copy Detecton (VMCD nthe followng) s even hgher. Fg. 3. Auto dual-threshold method to elmnate redundant vdeo frames, select keyframes. C 1f denotes the frst frame of Segment 1, C 1l the last frame of the Segment 1; C 2f denotes the frst frame of the Segment 2 IV. SVD-SIFT MATCHING In ths secton we dscuss the use of the SIFT descrptor n the SVD-matchng algorthm. As mentoned n the prevous secton SVD-matchng presented n [16] does not perform well when the baselne starts to ncrease. The reason for ths behavor s n the feature descrptor adopted. The orgnal algorthm uses the grey level values n a neghborhood of the keypont. As ponted out n Secton 2 ths descrpton s too senstve to changes n the vew-pont and more robust descrptor have been ntroduced so far. A comparatve study of the performance of varous feature descrptors showed that the SIFT descrptor s more robust than others wth respect to rotaton, scale changes, vew-pont change, and local affne transformatons. www.jcst.com 865
Fg. 4. Examples of features extracted. The ellpse around the feature ponts represents the support area of the feature. In the same work, cross-correlaton between the mage grey levels returned unstable performance, dependng on the knd of transformaton consdered. The consderatons above suggested the use of a SIFT descrptor, nstead of grey levels. The descrptor s assocated to scale and affne nvarant nterest ponts [27], brefly sketched n Secton 2. Some examples of such key ponts are shown n Fg.4. IV.GRAPH METHOD FOR VIDEO SIMILARITY CHECKING The graph-based vdeo sequence matchng method for vdeo copy detecton. The method s presented as follows: Step 1: Segment the vdeo frames and extract features of the key frames. Accordng to the method descrbed n Secton 3, we perform the dual-threshold method to segment the vdeo sequences, and then extract SIFT features of the key frames. Step 2: Match the query vdeo and target vdeo. Assume that Q c ¼ fc 1 Q ; C 2 Q ; C 3 Q ;... ; C m Q g and T c ¼ fc 1 T ; C 2 T ; C 3 T ;... ; C n T g are the segment sets of the query vdeo and target vdeo from Step 1, respectvely. For each C Q n the query vdeo, compute the smlarty sm(c Q ; C j T ), and return k largest matchng results. K ¼ _n, where n s the number of segments n the target set, and _ s set to 0.05 based on our emprcal study. Step 3: Generate the matchng result graph accordng to the matchng results. In the matchng result graph, the vertex M j represents a match between C Q and C j T. To determne whether there exsts an edge between two vertexes, two measures are evaluated. Tme drecton consstency: For M j and M lm, f there exsts (-1)*(j-m) then M j and M lm satsfy the tme drecton consstency. Tme jump degree: For M j and M lm, the tme jump degree between them s defned as t j lm max( t t j, t j t If the followng two condtons are satsfed, there exsts an edge between two vertexes: 1. The two vertexes should satsfy tme drecton consstency. 2. The tme jump degree t ( s a preset threshold based on our emprcal study). Condton 1 ndcates that f the query vdeo s a copy dervng from the target vdeo, then the vdeo subsequence temporal order between query vdeo and target vdeo must m ) be consstent, whch s reasonable n real applcaton. If Condton 1 s satsfed, Condton 2 s used to constran the tme span of two matchng results between the query vdeo and the target vdeo. If the tme span exceeds a certan threshold, t s consdered that there does not exst certan correlaton between the two matchng results. Ths method s smlar to the probablty model n [14].Also, as an example, the matchng results n can be converted nto a matchng result graph. Obvously, the matchng result s a drected acyclc graph. In the graph, n Case 1, because of volatng the condton of tme drecton consstency, t does not exst an edge between M 2;29 and M 3;26. For Case 2, although t meets tme drecton consstency, the tme jump between M 4;30 and M 5;70 exceeds the threshold, so t also does not exst an edge between M 4;30 and M 5;70. For each vertex of the matchng result graph, t may have more than one path or no path. For example, for vertex M 1;29, M 1;76, M 2;76, t has not any path to other vertexes (or say the path s the vertex tself). Step 4: Search the longest path n the matchng result graph.the problem of searchng copy vdeo sequences s now converted nto a problem of searchng some longest paths n the matchng result graph. The dynamc programmng method s used n ths paper. The method can search the longest path between two arbtrary vertexes n the matchng result graph. These longest paths can determne not only the locaton of the vdeo copes but also the tme length of the vdeo copes. Step 5: Output the result of detecton. For each vertex of the matchng result graph, t has more than one path or no path. As n Fg. 6, for the vertexes M 1;29, M 1;76, and M 2;76, they have no path to other vertexes, or only have path to the vertex tself. For M 1;26, four paths are avalable. Accordngly, we need to combne these paths that overlap on tme. Then, we can get some dscrete paths from the matchng result graph; t s thus easy to detect more than one copy segments by usng ths method. For each path, we use (3) to compute the smlarty of the vdeo m smk M j k 1 sm( path) m log(1 m) where m s the number of vertexes of the path, M j s the Q T vertex n the path, sm M j smc, C j. Accordng to the start pont and end pont of the path, we can obtan the tme stamp of the two copes. V.TIMESTAMP TO BLOCK THE MISBHEVING USERS Tmestamp A tmestamp s the tme at whch an event s recorded by a computer, not the tme of the event tself. In many cases, the dfference may be nconsequental: the tme at whch an event s recorded by a tmestamp (e.g., entered nto a log fle) should be close to the tme of the event. Ths data s usually presented n a consstent format, allowng for easy comparson of two dfferent records and trackng progress over tme; the practce of recordng tmestamps n a consstent manner along wth the actual www.jcst.com 866
data s called tme stampng. The sequental numberng of events s sometmes called tme stampng. Tmestamps are typcally used for loggng events or n a sequence of events (SOE), n whch case each event n the log or SOE s marked wth a tmestamp. In flesystems, tmestamp may mean the stored date/tme of creaton or modfcaton of a fle. TmeStamp for Vdeo copy detecton Server montors each and every users query vdeos. It wll contnuously montor the query vdeos of each and every user s communcaton. When montorng the server wll dentfy the coped frames from the nput queres.e. the matchng frame result s been verfed from the database whch s already been traned. It dentfes the unauthorzed user and block that dentfed user at partcular tme. I.e. based on tmestamp, agan the same user queres the same vdeo then that partcular user tmng wll ncreases (.e. 2 n to prevous block tme). Here we addtonally provde opportunty to all users.e. (user s tres to logn 3 attempts). Suppose f attempt of user s crossed the lmt then that partcular user wll dscard from the network. Advantages of Tmestamp Server wll dentfy msbehavng user n the proposed system that s based on tmestamp and block that dentfed user at partcular tme. Addtonally gve opportunty to all user s.e. user s tres to logn 3 attempts and block that dentfed user at partcular tme. VI.EXPERIMENTS Feature Extracton for Vdeo Copy Detecton In vdeo copy detecton, the sgnature s requred to be compact and effcent wth respectto large database. Besdes, the sgnature s also desred to be robust to varous codng varatons. In order to acheve ths goal,many sgnature and featureextracton methods are presented for the vdeo dentfcaton and copy detecton tasks[11] [12] [13] [14] [15] [16].As one of the common vsual features, color hstogram s extensvely used n vdeoretreval and dentfcaton [12] [11]. [12] apples compressed doman color features to form compact sgnature for fast vdeo search. In [11], each ndvdual frame s represented by four 178-bn color hstograms n the HSV color space. Spatal nformaton s ncorporated by parttonng the mage nto four quadrants. Despte certan level of success n [12] and [11], the drawback s also obvous, e.g. color hstogram s fragle to color dstorton and t s neffcent to descrbe each ndvdual key frame usng a color hstogram as n [12]. Another type of feature whch s robust to color dstorton s the ordnal feature. Hampapur et al. [13] compared performance of usng ordnal feature, moton feature and color feature respectvely for vdeo sequence matchng. It was concluded that ordnal sgnature had the best performance. The robustness of ordnal feature was also proved n [14].As a matter of fact, many works such as [3] and [14] also ncorporate the combned feature n order to mprove the performance of retreval and dentfcaton. Generally, the selecton of ordnal feature and color feature as sgnature for copy detecton task s motvated by the followng reasons: (1) Compared wth computatonal cost features such as edges, texture or refned color hstograms whch also contan spatal nformaton (e.g. color coherent vector appled n [15]), they are nexpensve to acqure (2) Such features can form compact sgnatures [] and retan perceptual meanng.(3) Ordnal features are mmune to global changes n the qualty of the vdeo and also contan spatal nformaton, hence are a good complement to color features [14]. Ordnal Feature Descrpton In our approach, we apply Ordnal Pattern Dstrbuton (OPD) hstogram proposed n [13] as the ordnal feature. Dfferent from [26], the feature sze s further compressed n ths paper, by usng more compact representaton of I frames. Fgure 2 depcts the operatons of extractng such features from a group of frames. For each channel c =Y, Cb, Cr, the vdeo clp s represented by OPD hstograms as: H OPD c (,... h1, h2,..., hl hn )0 h 1and h 1 Here N= 4! = 24 s the dmenson of the hstogram, namely the number of possble patterns mentoned above. The total dmenson of the ordnal feature s 3 24=72. Color Feature For the color feature, we characterze the color nformaton of a GoF by usng the cumulatve color nformaton of all the sub-sampled I frames n t. For computatonal smplcty, Cumulatve Color Dstrbuton (CCD) s also estmated usng the DC coeffcents from the I frames. The cumulatve hstograms of each channel (c=y, Cb, Cr) can be defned as: bkm 1 1 CCD HC H ( ) j 1,... B M b k where H denotes the color hstogram descrbng an ndvdual I frame n the segment. M s the total number of I frames n the wndow and B s the color bn number. In ths paper, B = 24 (unform quantzaton). Hence, the total dmenson of the color feature s also 3 24=72, representng three color channels. CONCLUSION In ths research propose a framework for content-based copy detecton and vdeo smlarty detecton. The proposed framework Based on the analyss, we use local feature of SIFT to descrbe vdeo frames. Snce the number of SIFT ponts extracted from a vdeo s large, so the copy detecton usng SIFT features has hgh computatonal cost. Then, we use a dual-threshold method to elmnate redundant vdeo frames and use the SVD-based method to compute the smlarty of two SIFT feature pont sets. After that graph based vdeo sequence matchng method are utlzed for matchng the each frame from the vdeo sequence Thus, detectng the copy vdeo becomes fndng the longest path www.jcst.com 867
n the matchng result graph are obtaned. Suppose f the result of the frame s matched,.e the matchng frame result s been verfed from the database whch s already been traned. It dentfes the unauthorzed user and block that dentfed user at partcular tme. I.e. based on tmestamp, agan the same user queres the same vdeo then that partcular user tmng wll ncreases (.e. 2 n to prevous block tme). REFERENCES [1] X. Wu, C.-W. Ngo, A. Hauptmann, and H.-K. Tan, Real-Tme Near-Duplcate Elmnaton for Web Vdeo Search wth Content and Context, Feb. 2009. [2] O. Ku cu ktunc, M. Bastan, U. Gu du kbay, and O. Ulusoy, Vdeo Copy Detecton Usng Multple Vsual Cues and MPEG-7 Descrptors, 2010. [3] J. Law-To, C. L, and A. Joly, Vdeo Copy Detecton: A Comparatve Study, Proc. ACM Int l Conf. Image and Vdeo Retreval, pp. 371-378, July 2007. [4] X. Zhou, L. Chen, A. Bouguettaya, Y. Shu, X. Zhou, and J.A. Taylor, Adaptve Subspace Symbolzaton for Content-Based Vdeo Detecton, IEEE Trans. Knowledge and Data Eng., vol. 22, no. 10, pp. 1372-1387, Oct. 2010. [5] A. Joly, O. Busson, and C. Frelcot, Content-Based Copy Retreval Usng Dstorton-Based Probablstc Smlarty Search, IEEE Trans. Multmeda, vol. 9, no. 2, pp. 293-306, Feb. 2007. [6] H. Lu, H. Lu, and X. Xue, SVD-SIFT for Web Near-Duplcate Image Detecton, Proc. IEEE Int l Conf. Image Processng (ICIP 10), pp. 1445-1448, 2010. [7] TRECVID 2008 Fnal Lst of Transformatons, http://www-nlpr. nst.gov/projects/tv2008/actve/copy.detecton/fnal.cbcd. vdeo.transformatons.pdf, 2008. [8] N. Gul, J.M. Gonza lez-lnares, J.R. Co zar, and E.L. Zapata, A Clusterng Technque for Vdeo Copy Detecton, Proc. Thrd Iberan Conf. Pattern Recognton and Image Analyss, Part I, pp. 452-458, June 2007. [9] K. Sze, K. Lam, and G. Qu, A New Key Frame Representaton for Vdeo Segment Retreval, IEEE Trans. Crcuts and Systems Vdeo Technology, vol. 15, no. 9, pp. 1148-1155, Sept. 2005. [10] J. Yuan, L.-Y. Duan, Q. Tan, S. Ranganath, and C. Xu, Fast and Robust Short Vdeo Clp Search for Copy Detecton, Proc. Pacfc Rm Conf. Multmeda (PCM), 2004. [11] S. Cheung and A. Zakhor, Effcent vdeo smlarty measurement wth vdeo sgnature, In IEEE Trans. on Crcuts and System for Vdeo Technology, vol. 13, pp. 59-74, 2003. [12] M.R. Naphade et al., A Novel Scheme for Fast and Effcent Vdeo Sequence Matchng Usng Compact Sgnatures, In Proc. SPIE, Storage and Retreval for Meda Databases 2000, Vol. 3972, pp. 564-572, 2000. [13] A. Hampapur, K. Hyun, and R. Bolle., Comparson of Sequence Matchng Technques for Vdeo Copy Detecton, In SPIE. Storage and Retreval for Meda Databases 2002, vol. 4676, pp. 194-201, San Jose, CA, USA, Jan. 2002. [14] Junsong Yuan et al., Fast and Robust Search Method for Short Vdeo Clps from Large Vdeo Collecton, n Proc. of ICPR 04, Cambrdge, UK, Aug. 2004. [15] R. Lenhart et al., VsualGREP: A Systematc method to compare and retreve vdeo sequences, InSPIE. Storage and Retreval fro Image and Vdeo Database VI, Vo. 3312, 1998. [16] J. Oostveen et al., Feature extracton and a database strategy for vdeo fngerprntng, In Vsual 2002, LNCS 2314, pp. 117-128, 2002. www.jcst.com 868