756 JOURAL OF COPUTERS, VOL. 8, O., OVEBER 03 Vdeo Watermarkng Algorthm Based on Relatve Relatonshp of DCT Coeffcents Cheng ngzh, Du Yanpng, Wang Yan Bejng Insttute of Graphc Communcaton, Bejng, Chna Emal: chengmz@bgc.edu.cn, duyanpng@bgc.edu.cn, wangyanz@bgc.edu.cn X nchao, Yuan Kaguo School of Computng, Bejng Unversty of Posts and Telecommuncatons, Bejng, Chna Emal: mnchao@gmal.com, flyngdreamng@gmal.com Abstract A vdeo watermarkng algorthm based on relatve relatonshp of DCT coeffcents s proposed n ths paper. By modfcaton of coeffcents on chosen postons, the robustness of watermark embeddng s ensured. eanwhle, as the relatonshp of coeffcents remans unchanged, the qualty of vdeo s not affected. The analyss results shows that the performance of proposed algorthm s better than the typcal algorthms. Index Terms Vdeo Watermark; Relatve Relatonshp; DCT; BCH; Arnold Transformaton I. ITRODUCTIO In recent decade, as mult-meda nformaton become more and more dgtalzed, technology of vdeo dgtalzaton becomes rapdly advanced n way of nformaton nqury, presentaton, storage and transfer. However, ssues have been rased smultaneously wth the most phenomenon one as ntellectual property theft. The cause of ths ssue should be prmarly blamed on the low lmtaton of duplcaton. Watermarkng technology become more often used n the area of copy protecton, wth ths method, orgnal vdeo owner could nsert the property dentfcaton nformaton nto the vdeo. The vdeo fle could stll be used or transferred normally. The orgnalty of vdeo s always traceable, ntellectual property s as a result beng protected. Although watermarkng technology cannot prevent llegal duplcaton of vdeo data, authors and publshers could stll track the source of unauthorzed vdeo, and stop the llegal behavor as soon as possble. Smlar wth nformaton hdng technques, watermark algorthms work by embeddng a encodng sgnal nto the vdeo fle. The sgnal could generally nclude the owner dentfcaton, publshng dates and seral number. All the nformaton s used to track the status of publshng and usage of protected vdeos [] []. Vdeo watermarks can be ether vsble or nvsble. The logo of authors, photocopy of lbrary books etc. always showed as the vsble watermark. Invsble watermarks n the other way, does not affect the vdeo appearance, copy nformaton s coded nsde wthout other nfluence. Accordng to the preservaton of vdeo qualty, nvsble watermarks are more commonly used []. Vdeo watermarkng algorthms could also be classfed nto two categores: watermark of spatal doman and watermark of frequency doman. Spatal doman watermarkng algorthms embed the watermark by drectly modfyng the pxel values of the host vdeo, whle frequency doman watermarkng algorthms embed watermark by modulatng the frequency coeffcents. The latter one s more robust when sufferng ordnary vdeo manpulatons. Furthermore, among the several frequency domans-dct, DWT, DFT, algorthms based on DCT doman are more capablty n dgtal vdeo codec standards, so t draws majorty attentons of researchers [3]. In ths paper, the algorthm we are presentng modfes coeffcents wthout changng ther relatve relatonshp, whle the former algorthms tends to change coeffcents arbtrarly, causng problem n the qualty of vdeo. In fact, vdeo qualty could be even better, as the characterstcs of the mage data have been taken advantage. The man contrbutons of ths paper are as follows. We reveal the prncple of common watermarkng algorthms: storng bnary nformaton n the host data usng bnary relatonshp. And we proposed a new vdeo watermarkng algorthm storng nformaton by a dfferent bnary relatonshp whch mproves the vdeo qualty. The rest of ths paper s organzed as follows. Defntons of DCT and BCH code are ntroduced n secton II. Prncple and procedures of the proposed watermark algorthm are dscussed n detal n secton III. In secton IV, performances of our algorthm are evaluated and analyzed. Fnally, conclusons and future works are rased n secton V. II. RELATED WORK A. DCT and Vdeo Codec Standards DCT, short for Dscrete Cosne Transform, s wdely used n popular vdeo compresson standards, such as 03 ACADEY PUBLISHER do:0.4304/jcp.8..756-76
JOURAL OF COPUTERS, VOL. 8, O., OVEBER 03 757 H.64 and PEG4. The purpose of DCT transform n vdeo codec s to convert mage or moton-compensated resdual data nto another doman, the frequency, or more commonly called, transform doman. The choce of transform depends on a number of crtera: ) Data n the transform doman should be decorrelated,.e. separated nto components wthn mnmal nter-dependence, and compact,.e. most of the energy n the transformed data should be concentrated nto a small number of values. ) The transform should be reversble. 3) The transform should be computatonally tractable, e.g. less memory requrement, achevable usng lmtedprecson arthmetc, fewer arthmetc operatons, etc. There are plenty of vdeo compresson algorthms and the most popular transforms tend to fall nto two categores, block-based and mage-based. Examples of block-based transforms nclude the KLT (Karhunen- Loeve Transform), SVD (Sngular Value Decomposton), and the ever-popular DCT (Dscrete Cosne Transform). Each of these operates on blocks of * mage or resdual samples and hence the mage s processed n unts of a block. Block transforms have less memory requrements and are well suted to compresson of blockbased moton compensaton resduals whle sufferng from artfacts at block edges, however. Image-based transforms operate on an entre mage or frame or a large secton of the mage known as a tle. The most popular mage transform s the DWT (Dscrete Wavelet Transform), or wavelet. Image transforms such as the DWT have been shown to out-perform block transforms for stll mage compresson but they tend to have hgher memory requrements because the whole mage or tle s processed as a unt and they do not necessarly ft well wth block-based moton compresson whch s the most mportant and effectve operaton of vdeo compresson. The approxmatons to the DCT are featured n H.64 for ts hgh effcency and consderable ablty of decorrelaton. The Dscrete Cosne Transform operates on X, a block of * samples, typcally mage samples or resdual values after predcton, to create Y, an * block of coeffcents. The acton of the forward DCT of a sample block s gven by [4]: ( ) ( ) ( ) ( ) ( ) ( ) + uπ j + vπ Y u, v = a u a v X, j cos cos () a = 0 j= 0, u = 0, u =,,..., ( u) =, a() v =, v = 0, v =,,..., Where Y stands for the DCT coeffcent matrx of X, n whch Y ( 0,0) s called DC coeffcent, and others are mentoned as AC coeffcents. The nverse DCT of a sample block s gven by: ( ) ( ) ( ) ( ) ( ) ( ) + uπ j + vπ (3) X, j = a u a v Y u, v cos cos u= 0 v= 0 = 0,,,...,, j = 0,,,...,. () B. BCH Code In codng theory, BCH code s one of the cyclc errorcorrectng codes that are constructed usng fnte felds. BCH code s nvented n 959 by Hocquenghem, Bose, Chaudhur ndependently. One of the key features of BCH codes s that durng code desgn, there s a precse control over the number of symbol errors correctable by the code. In partcular, t s possble to desgn bnary BCH codes that can correct multple bt errors. Another advantage of BCH codes s the ease of use as they can be decoded, namely, va an algebrac method known as syndrome decodng. Ths smplfes the desgn of the decoder for these codes, usng small low-power electronc hardware. Wth all these advantages, BCH code s wdely used n modern communcaton systems. Gven t as the number of errors desgns to correct, ( x) m be the mnmal polynomal, d stands for the code mnmal dstance, where d d 0 = t +, BCH code has the length of n = m, whose generator polynomal s [5]: g( x) = LC [ m ( x), m3 ( x)... mt ( x) ] (4) III. PROPOSED ETHOD A. Bnary Relatve Relatonshp The storage of watermarkng nformaton n vdeo frames usually depends on the relatonshps of coeffcents, and most of whch are bnary relatonshp. Ether watermarkng algorthms based on energy modulaton [4], or the ones depend on correlaton detecton and spread spectrum, the most basc prncple could be expressed as followng [][6-0]. As the basc nformaton unt n dgtal storage, bnary relatonshp s the most basc relatonshp, as the formula below: W = R ( x, x ) (5) Where x, x are the values based on mage contents, or const values whch are pre-defned. For nstance, a knd of watermarkng algorthm based on energy modulaton s proposed n paper [4]. In whch the watermark nformaton are kept on the bnary relatonshp of the coeffcent energy n each block and the average energy of the whole mage. Ths could be represented as ths formula, whch has represented a knd of bnary relatonshp:, f P P W = R( P, P ) = (6) 0, f P < P Where W stands for the th bt of watermark nformaton, P expresses the energy of th block, and P s the average energy of blocks of the whole mage. Another knd of example algorthms comes from artcles [][6-0], whch proposed watermark algorthms based on spread spectrum and correlaton detecton. Wthn these algorthms, watermark nformaton are stored by ths formula, whch also represents a knd of bnary relatonshp: 03 ACADEY PUBLISHER
758 JOURAL OF COPUTERS, VOL. 8, O., OVEBER 03, f Corr( S, Sw) Threshold W = R( S, S ) = (7) w 0, f Corr( S, Sw) < Threshold Where S stands for the coeffcent sequence, S s the w detecton sequence generated by watermark nformaton. Based on the prncples mentoned before, a watermark algorthm based on bnary relatonshp of three coeffcents s proposed, whch stores watermark nformaton by modulatng the dstances between coeffcents. Compared wth other algorthms whch modfed the coeffcents arbtrarly, ths algorthm changes the coeffcents when the relatonshp of whch s kept, and takes advantage of the characterstcs of mage data, whch could mproves the vdeo qualty. The proposed algorthm could be expressed n the followng formula: 0, f D D W = R( D, D ) = (8), f D > D D md Coef D Coef md It should be emphaszed that Coef, Coef, and md Coef are not coeffcents whose postons are fxed. They are a ordered coeffcent sequence from three predefned postons, where Coef Coefmd Coef. The proposed algorthm keeps the relatonshp of three coeffcents unchanged whle embeddng the watermarks, whch means the characterstc of vdeo mage data are sustaned, and excess modfcatons are avod. B. Watermark Embeddng Gven a vdeo of sze *. The vdeo frame are dvded nto blocks of 8*8, whch are expressed as D ( x, y), x, y =,,3,..., 8, =,,3..., L, L = ( / 8 * / 8 ). For each block, a block of DCT coeffcents s generated, marked as C ( u, v), u, v =,,3,...,8, =,,3..., L. To generate watermark nformaton, a H * H bnary btmap s converted to bt stream W. Then, encode W wth a BCH encoder by the parameters = 63, K = 8, a bt stream of W ' s generated. Embed W ' nto each 8*8 DCT coeffcent block C ( u, v) bt by bt usng the followng steps. a) Select three coeffcents from C, sort them nto Coef Coef md Coef. b) Calculate the dstance between Coef and Coef, D total Coef. To ensure the robustness of algorthm, a threshold Threshold s defned to ndcate the mnmal dstance of Coef and Coef. If D total < Threshold, then the algorthm modfes the values of Coef and Coef, to ncrease the dstance of them to Threshold, as followng: Threshold Dtotal Threshold Dtotal + (9) If D total Threshold, then wll not be modfed, that s : c) Embed W ' nto C to get C', where W ' { 0,}. If W '= 0, then do the followng operaton: Dtotal md = + (0) Scale If W '=, then do the followng operaton: Dtotal md = () Scale Where Scale 3 s a predefned const. Ths const number stands for the dstance of and md +. The bgger the value s, the stronger the watermark sgnals are embedded. d) Convert each embedded coeffcent block C' to spatal doman- a embedded pxel block D'. Compose all the embedded blocks nto a vdeo frame I '. C. Watermark Detecton Just lke the embeddng procedure, gven a suspcous vdeo of sze *. The vdeo frame are dvded nto 8*8 blocks D' ( x, y), x, y =,,3,..., 8, =,,3..., L, L = ( / 8 * /8 ). For each block, a block of DCT coeffcents s generated, marked as C' ( u, v), u, v =,,3,...,8, =,,3..., L. Detect watermark nformaton WE ' from each 8*8 DCT coeffcent block C' ( u, v) bt by bt usng the followng steps. a) Select the three coeffcents used n the embeddng procedure. Then sort them to generate the coeffcent sequence md. b) Calculate D = md Coef ' and D Coef, then each bt of WE ' s md generated. If D D, then WE ' = 0. If D > D, then WE ' =. () To generate watermark nformaton, a H * H bnary btmap s converted to bt stream W. Then, encode W wth a BCH encoder by the parameters = 63, K = 8, a bt stream of W ' s generated. Decode WE ' wth a BCH decoder by the parameters = 63, K = 8, a bt stream WE s generated. Then, WE s converted to a H * H bnary btmap, as the extracted watermark mage. 03 ACADEY PUBLISHER
JOURAL OF COPUTERS, VOL. 8, O., OVEBER 03 759 IV. A. Algorthm Parameters EXPERIETAL RESULTS Parameters Threshold = 40 and Scale = 6 are selected to evaluated the performances of the algorthm. Four test vdeos are chosen to evaluate the algorthm: flower (35*88), foreman (35*88), mother (35*88), waterfall (35*88). A 6*6 bnary btmap s used as the watermark mage, as followng: B. Embed Performances Some of the embedded vdeo frames are showed n Fgure to Fgure 8. Fgure 4. Watermarked frames n foreman_cf.yuv. PSR= 38.5698. Fgure. Orgnal frames n flower_cf.yuv. PSR= 35.6507. Fgure 5. Orgnal frames n mother_cf.yuv. PSR= 39.009. Fgure. Watermarked frames n flower_cf.yuv. PSR= 35.6507. Fgure 6. Watermarked frames n mother_cf.yuv. PSR= 39.009. Fgure 3. Orgnal frames n foreman_cf.yuv. PSR= 38.5698. Fgure 7. Orgnal frames n waterfall_cf.yuv. PSR= 39.6570. 03 ACADEY PUBLISHER
760 JOURAL OF COPUTERS, VOL. 8, O., OVEBER 03 Fgure 8. Watermarked frames n waterfall_cf.yuv. PSR= 39.6570. C. Extract Performances Ths paper evaluates the algorthm robustness aganst JPEG compresson (Wth the compresson strengths of 35, 40, 50, 60, 70, 80, 90, 00) and vdeo zoomng attacks (Wth the scales of 50%, 75%, 90%, 0%, 50%, 00%). The PSR between embedded vdeo frames and orgnal frames are calculated, and C values between extracted watermark mage and orgnal one are also evaluated. PSR(Peak Sgnal to ose Rato) s used to evaluate the vsual qualty after watermark embeddng, larger value means the better vsual qualty. Defnton s lsted below: 55 (3) PSR = 0*lg ( I(, j) I '(, j) ) * = j= Where I s the orgnal vdeo frame, and I ' means embedded frame. C (ormalzed utual Correlaton Coeffcent) s the standard to evaluate the smlarty between extracted watermark mage and orgnal one, the larger the better. C = L = L = W W () WE() () * WE() L = The evaluate results are gven n table I. TABLE I. EXTRACTED WATERARK RESULTS flower foreman mother waterfall PSR 35.6507 38.5698 39.009 39.6570 JPEG_30 0.96 0.833 0.7334 0.9896 JPEG_40 0.985 0.980 0.876.0000 JPEG_50.0000.0000.0000.0000 JPEG_60.0000.0000.0000.0000 JPEG_70.0000.0000.0000.0000 JPEG_80.0000.0000.0000.0000 JPEG_90.0000.0000.0000.0000 JPEG_00.0000.0000.0000.0000 RESC_50.0000.0000.0000.0000 RESC_75.0000.0000.0000.0000 RESC_90 0.985 0.9463.0000 0.8977 RESC_0.0000 0.936.0000 0.835 RESC_50 0.9789 0.909.0000 0.7587 RESC_00 0.9789 0.948.0000 0.9547 D. Algorthm Comparson In [], a typcal DCT watermark algorthm mentoned before s evaluated. Wth the proposed algorthm, results are showed n Fgure 9 and Fgure 0. The comparson results between the typcal algorthm and the proposed algorthm are gven n table II. Compared wth the typcal algorthm, the robust of proposed algorthm aganst JEPG compresson and rescalng s mproved, and the vsual qualty of embedded vdeo s better. (4) 03 ACADEY PUBLISHER
JOURAL OF COPUTERS, VOL. 8, O., OVEBER 03 76 o. K0005007, K0005006 and KZ000505, BIGC Project o. E-a-0-5, E-6-0-8, E-a-0-7 and E-b-03-6, Bejng Unversty of Posts and Telecommuncatons under Grant 0RC07. REFERECES Fgure 9. Orgnal mage of Lena.bmp. PSR= 38.8547 Fgure 0. Watermarked mage of Lena.bmp. PSR= 38.8547. TABLE II. COPARISO OF THE PROPOSED ETHOD WITH THE TYPICAL ALGORITH Lena Proposed Alg. Typcal Alg. PSR 38.8547 35.76 JPEG_30 0.8565 0.779 JPEG_40 0.903 0.95 JPEG_50 0.9967 0.9868 V. COCLUSIO In ths paper, a vdeo watermark algorthm n DCT doman based on bnary relatonshp of three coeffcents s proposed, whch stores watermark nformaton by modulatng the dstances between coeffcents. The processes of ths watermarkng scheme, ncludng embeddng detecton, are descrbed n detal. The evaluaton has proved that our method can survve compresson and rescale attacks, and further mprovements could be made to get better performance. Improvements could be made from followng aspects: ) Increasng the number of embeddng postons. ) Scramblng the watermark mage before embeddng to mprove the robustness. 3) Evaluatng the robustness of coeffcents n dfferent postons. ACKOWLEDGET Ths work was supported n part by a grant from Bejng atural Scence Foundaton under Grant o. 406, Scentfc Research Common Program of Bejng uncpal Commsson of Educaton under Grant [] Xangpng eng, Yan Gao. Electrc Systems Analyss []. Bejng: Hgher Educaton Press, 004: 3-. [] Pk Wah Chan, chael R. Lyu, and Roland T. Chn. A ovel Scheme for Hybrd Dgtal Vdeo Watermarkng: Approach, Evaluaton and Expermentaton [J]. IEEE Transactons on Crcuts and Systems for Vdeo Technology, VOL. 5, O., December 005: 638-649. [3] Alexa Brassoul, Panagots Tsakaldes, and Athanasos Stourats. Hdden essages n Heavy-Tals: DCT- Doman Watermark Detecton Usng Alpha-Stable odels [J]. IEEE Transactons on ultmeda. VOL. 7, O. 4, August 005: 700-75. [4] Gerhard C. Langelaar, Iwan Setyawan, and Regnald L. Lagendjk. Watermakng Dgtal Image and Vdeo Data [J]. IEEE Sgnal Processng agazne. September 000: 0-46. [5] Jongpan Zhou, Qnhua Pang,Dawo Xu,Welng Wu. Communcaton Theory []. BUPT Press,00: 6. [6] Zhou Xnmn, Yu Janpng. Attack odel and Performance Evaluaton of Text Dgtal Watermarkng. Journal of Computers, 00, (5): 933-94. [7] J.J.K. Ó Ruanadh and T. Pun, Rotaton, scale and translaton nvarant spread spectrum dgtal mage watermarkng [J], Sgnal Processng, vol. 66, ay 998, no.3, pp. 303-37. [8] A. Herrgel, H. Petersen, J. Ó Ruanadh, T. Pun, and P. Shelby, Copy technques for dgtal mages based on asymmetrc cryptographc technques [C], presented at Workshop on Informaton Hdng, Portland, Oregon, USA, Apr. 998. [9] Deng nghu, Zeng Qngshuang, Zhou Xul. A Robust Watermarkng Aganst Shearng Based on Improved S- Radon Transformaton. Journal of Computers, 0, 0(7): 549-556. [0] S. Perera, J.J.K. Ó Ruanadh, F. Degullaume, G. Csurka, and T. Pun, Template based recovery of Fourer-based watermarks usng log-polar and log-log maps [C], n Proc. IEEE ultmeda Systems 99, Int. Conf. ultmeda Computng and Systems, Florence, Italy, June 7-, 999.. [] Rao Zhjan, Chang Janpng, Adaptve Spread Spectrum Watermarkng Algorthem Based on DCT blocks, Technology Square, 005.5: 9-4. ngzh Cheng was born n 974 n Hube Chna. He receved hs B.S. degree n Communcaton Engneerng from Chongqng Unversty of Posts and Telecommuncatons, Chongqng, Chna, n 996, and receved hs Ph.D. degree n Cryptography from Bejng Unversty of Posts and Telecommuncatons, Bejng, Chna, n 00. From 00, Dr. Cheng s workng n Packagng Antcounterfetng and Dgtal Content Securty n Bejng Insttute of Graphc Communcaton as a lecturer. 03 ACADEY PUBLISHER
76 JOURAL OF COPUTERS, VOL. 8, O., OVEBER 03 Yanpng Du was born n 97 n Shanx Chna. She receved her B.S. degree n computer scence and technology from Shanx Unversty, Shanx, Chna, n 994, and receved her Ph.D. degree n control theory and engneerng from Bejng Jaotong Unversty, Bejng, Chna, n 005. From 006, Dr. Du s workng n Packagng Antcounterfetng and Dgtal Content Securty n Bejng Insttute of Graphc Communcaton as a professor. Yan Wang was born n 975 n Hube Chna. She receved her B.S. degree n hgh voltage technology from Huazhong Unversty of Scence and Technology, Hube, Chna, n 996, and receved her Ph.D. degree n weapon system and applcaton engneerng from Bejng Insttute of Technology, Bejng, Chna, n 005. Her research topc s electronc nformaton and sgnal processng. From 006, Dr. Wang s workng n Bejng Insttute of Graphc Communcaton as a lecturer. nchao X was born 987 n Shanx Chna. He receved hs B.S. degree n computer scence and technology from Shanx Unversty, Shanx, Chna, n 009, and receved hs Ph.D. degree n computer scence and technology from Bejng Unversty of Posts and Telecommuncatons, Bejng, Chna, n 03. Dr. X s workng n Image and Vdeo Watermarkng Theory as a lecturer n Bejng Unversty of Posts and Telecommuncatons. Kaguo Yuan was born n 98 n Guzhou Chna. He receved hs Ph.D degree n sngal and nformaton processng from Bejng Unversty of Posts and Telecommuncatons, Bejng, Chna, n 009. Dr. Yuan s workng n Informaton Securty as a lecturer n Bejng Unversty of Posts and Telecommuncatons. 03 ACADEY PUBLISHER