An Efficient Blind Watermarking Method based on Significant Difference of Wavelet Tree Quantization using Adaptive Threshold

Transcription

1 Internatonal Journal of Electroncs and Electrcal Engneerng Vol., No., June 3 An Effcent Blnd ng Method based on Sgnfcant Dfference of Wavelet Tree Quantzaton usng Adaptve Threshold Then Huynh The Unversty of Techncal Educaton Ho Ch Mnh Cty, HCM Cty, Vetnam Emal: thethen8@gmal.com Thuong Le Ten Ho Ch Mnh Cty Unversty of Technology, HCM Cty, Vetnam Emal: thuongle@hcmut.edu.vn qualty. But ths method s dsadvantage s that t cannot effectvely resst low-pass flterng attacks such as medan flterng and Gaussan flterng. The authors n [3] proposed the method based on sgnfcant dfference quantzaton technque; that s, every seven-wavelet coeffcents n 3-level DWT sub-band was grouped as a block and the watermark bt s embedded by quantzng the dfference between two largest coeffcents. Based on nsgnfcant coeffcents of a wavelet tree, the authors proposed the method that mproved the qualty of the watermarked mage, but the qualty of recovery watermark was decreased. An ntellgent watermarkng method based on partcle swarm optmzaton (PSO) n [5] s the mproved method n [3]. Ths method has been resolved the conflct between mperceptblty and robustness of watermarkng by selectng randomly coeffcents to make up a block to avod the watermark beng detected and the strength of embeddng watermark among dfferent blocks s adaptve by PSO method to balance the robustness and perceptual transparency. By scalng the magntude of the sgnfcant dfference between the two largest wavelet coeffcents n a wavelet tree, the authors [6] mproved the robustness of the watermarkng. However, extracted watermark cannot ensure the qualty under JPEG compresson. Embeddng the nformaton nto the sgnfcant wavelet coeffcent n those dynamc blocks wth strong edges accordng to a "Bnary Algorthm" was proposed n research [7]. Snce the change to the blocks wth strong edge strengths s less vsble to human eyes. The threshold n extracton process was lmted by the demanded sze of the watermark. The authors n [8] used wavelet packet decomposton technque for both the orgnal mage and the watermark by embeddng watermark nto low-frequency sub-band and comparng the correlaton between recovery and orgnal watermark to decde the exstence of watermark. The authors n [9] suggested a method that usng statstcal characterstcs of coeffcents to quantze. Ths method used two quantzaton algorthm for two phases of watermarkng system. Usng the same method, the Abstract In ths research, we ntroduce an effcent blnd watermarkng method for gray mage based on wavelet tree quantzaton usng adaptve threshold n extracton process. Based on dfference value between two largest coeffcents, each scrambled bnary watermark bt s embedded nto each block whch created by four LH3 coeffcents and one LH4 coeffcent. In extracton process, we compare dfference value n each block to adaptve threshold to recover watermark bt. The qualty of extracted watermark depends on the threshold whch s determned by the Weghted Wthn-Class Varance algorthm. Performance of proposed method s represented through expermental results under varous attacks such as, Hstogram Equalzaton, Croppng, Low-pass Flterng, Gaussan nose, Salt & Pepper nose and JPEG compresson. Index Terms blnd watermarkng, 3-level DWT, sgnfcant dfference, adaptve threshold, wthn-class varance I. INTRODUCTION Due to the advantage of the Internet, people can arbtrarly and access or dstrbute dgtal products. Hence, the dgtal watermarkng technque has been used n multmeda products to protect copyrght and authentcate mage... Ths technque embeds nformaton nto the dgtal contents so that the vewer cannot see any nformaton. However, there are many problems n the watermarkng system: the watermark has to ensure not degradng the qualty of the cover mage and beng perceptually nvsble for human eyes; the watermark must be robust to resst dfferent attacks and the watermarkng system s blndness. Several wavelet tree-based watermarkng methods are proposed based on the qualfed sgnfcant wavelet tree (QSWT) [,, 4, 5, and 6]. In [], the authors embedded watermark n each of two sub-bands of wavelet tree. The authors n [] mproved the method n [] by usng four trees to represent two watermark bts to enhance vsual Manuscrpt receved February, 3; revsed May, 3. do:.7/jeee

2 Internatonal Journal of Electroncs and Electrcal Engneerng Vol., No., June 3 authors n [] embedded watermark by modfyng the center coeffcent n each block and the value of ths coeffcent depend on the watermark bt and the mean of four cross neghbor coeffcents. Another way to group was represented n []: LH3 and HL3 sub-band are dvded nto equal four sectons. Each block was grouped by one coeffcent n each secton. Before that, the basc method s proposed n [] embeddng each watermark bt by comparng the dfference of two largest sgnfcant coeffcents n each wavelet tree to the average value. The maxmum coeffcent was modfed based on value of watermark bt. In [3], the authors embedded each pxel of watermark nto the wavelet coeffcent of mddle and low frequency of a block n the cover mage. However, the proposed method s non-blnd watermarkng system; that s, usng orgnal mage for extracton process. Addtons, the authors n [4] used a re-watermarkng approach employng blnd, quantzaton-based and robust technque to embed multple fngerprntng. They nvestgated two extensons to a wavelet coeffcent-tree based embeddng technque that turn out to mprove detecton performance. In ths research, an effcent blnd watermarkng method for mage based on the sgnfcant dfference of wavelet coeffcents s developed from [3]. Due to the mportance of determnng threshold value n extracton process, ths value could be determned based on the probablty densty functon (pdf) of blocks along dfference values nstead of proposed algorthm n [3]. By computng the weghed wthn-class varance [5], we separate pdf plot nto two parts and the dfference value chosen as threshold has mnmum varance. Smulaton results under dfferent attack cases demonstrate method's performance through PSNR of embedded mage and NC of extracted watermark. We have sze of HL3 after decomposng s when usng mage. There are watermark bts are embedded nto blocks, whch are grouped based on wavelet tree as Fg. ; that s, one coeffcent of HL4 and four coeffcents of the same orentaton n the same locaton n HL3 are grouped as a block. Each wavelet tree conssts of coeffcents for a wavelet tree constructed at each node from HL4 to HL3. Each 4-level sub-band, we have 4 coeffcents whch s the maxmum number of watermark can be embedded. However, due to watermark wth sze, we only need 5 blocks n embedded and extracton algorthm. For embeddng process, we modfy the maxmum coeffcent by the fxed quantty, called quantzaton value, based on the sgnfcant dfference of two largest coeffcents n each block and value of watermark bt. Ths value s the mean of all sgnfcant dfference values of all mage used. For one mage, we can fnd ths value as follows: N (max sec ) N () where N: number of blocks, max and sec are two th largest coeffcents n have: block. For many mages, we T m j....() m j where m: number of mages. The mages are consdered n smulaton nclude Lena, Baboon, Pepper, Goldhll and Salboat. The value of each mage and average of them s represented n TABLE I. The quantzaton value used n ths paper s the mean of all above mages. II. THE PROPOSED METHOD TABLE I. QUANTIZATION VALUE OF EACH IMAGE A. The ng Method ng model s represented n Fg.. The watermark after scramblng wth securty key wll be hd nto the host mage. We decompose ths mage at 4-level DWT to four sub-bands. Based on the wavelet tree technque, three are three super wavelet trees: HH3-HH3HH-HH s easly elmnated when usng the JPEG lossy compresson. HL4-HL3-HL-HL s sutable to embed the watermark n ths paper to satsfy extractng watermark under dfferent attacks nstead of LH sub-band famly. To lmt computng capacty, we only choose HL4 and HL3 for embeddng process. Lena ℰ 5 Baboon Pepper Goldhll Salboat T=Average 58 4 B. The Embeddng Process Comparng sgnfcant dfference to threshold to extract watermark bt Coeffcent of the HL LH3 LH HL3 HH3 LH HL HH Extracted HL HH Extracted Fgure. Embeddng watermark based on wavelet tree quantzaton + Embeddng Algorthm Host + Key Channel + Extracton Algorthm The watermark mage used to embed nto the host gray mage has sze 5 bts wth value or. After scramblng watermark wth securty number, each bt wll be embedded nto each block, respectvely. Each block have fve coeffcents (4 HL3 coeffcents + HL4 coeffcent) s created after decomposng Scambleb bt Attacks Fgure. ng model. 99

3 Internatonal Journal of Electroncs and Electrcal Engneerng Vol., No., June 3 gray mage. In each block, we seek two values: the largest coeffcent and sec - the second largest coeffcent sgnfcant and (max sec ). For th dfference value. Wthout attacks, we can determne the threshold value to extract watermark exactly. However, the change of pdf s very lght under some attacks, such as, hstogram equalzaton, Gaussan flter; otherwse, some attacks affect pdf extreme serously, such as average flter. In these cases, we cannot extract correctly encoded watermark; t can be explaned that the watermark bt are embedded, sgnfcant dfferences wll be ; however under varous attacks, they are larger than value. max - dfference bt of watermark, we perform the embeddng algorthm as the followng steps for block: Wth watermark bt : max where max ;f ( T ) max ;f ( T ) (3) T Sgnfcant dfference () Modfed sgnfcant () Weghted wthnof each block dfference value class varance Compute average Seek the value mnmum value Wth watermark bt : max max / sec sec / Fgure 5. The flowchart for determnng threshold The qualty of embedded mage depend on value T n the case we use same watermark. The detaled of embeddng process can be showed n Fg blocks 3-level DWT HL3 subband Embeddng Algorthm In ths paper, we proposed an algorthm to determne adaptve threshold based on computng the weghed wthn-class varance. In more detal, we fnd the threshold that gets mnmzed the weghted wthn-class varance, or ths turn out to be the same as maxmzng the between-class varance. However, we need to modfy sgnfcant dfference of the blocks whch have. Ths preprocess does not affect the qualty of recovery watermark because these blocks wll extract exactly snce threshold value s always less than. The proposed algorthm for determnng threshold s showed n Fg. 5. The process () n Fg. 5 can be performed as follows: HL4 subband pseudo-random sequence È 4-level DWT Host Bnary watermark Scramble HL3 subband IDWT 4 blocks Restructure Fgure 3. The flowchart of embeddng process. C. The Extracton Process bt embedded nto each block can be recovered based on comparng sgnfcant dfference to threshold - denoted as. After embeddng, dfference n each block of cover mage wll be or larger than as (3) and (4). In wth non-attack case, watermark s extracted exactly when the threshold n the range ( y ). However, ths value can be changed n dfferent attack cases; that s, extracted bt s nstead of. When fnd out the threshold, we can extracted as below equaton: bt ;f ( d ) ;f ( d ) d ;f ( d ) ;f ( d ) d w (d ) W (d ) (d ) W (d ) (d ).(7) where class probabltes estmated: and d W (d ) p() (5) W (d ) max ( d ) p() =d+ Attack Type: Rotaton And the class means are gven by:.4 d.8 ( d ) PDF p( ) W (d ).. 3 Sgnfcant Dfference are (8) Non-attack. (6) The detaled process () n Fg. 5 can be calculated as:.5. PDF Determne adaptve threshold y Average value ℰ (4) 4 Fgure 4. Probablty densty functon 3 Sgnfcant Dfference x ( n) (d ) 4 of sgnfcant dfference max ( d ) d p() W (d ) of cover mage Fnally, the ndvdual class varances are: We can consder the probablty densty functon x(n) of sgnfcant dfference n Fg. 4. The functon (d ) ( (d )) x(n) represents the number of block havng the same d p() W (d ) (9) ()

4 Internatonal Journal of Electroncs and Electrcal Engneerng Vol., No., June 3 max( d ) ( d) ( ( d)) d p () W( d) These calculaton steps were done through the range of d max d and pck out the value that value [, ( )] mnmzed ( d) W sgnfcant dfference value whch has. The adaptve threshold s the ( mn). If W there are many values, the threshold wll be the mean of them: mean d d mn () ( / W ( ) ) After determnng adaptve threshold, we use (5) to extract watermark. III. EXPERIMENTAL RESULT We use the peak sgnal-to-nose rato PSNR to evaluate the qualty between embedded mage and orgnal mage. Ths formula s defned as follows: PSNR 55 log ( ) MSE () where MSE s computed as below equaton: M N MSE I j I j M N j (, ) (, ) (3) where M and N are heght and wdth of the mage, are the grey value respectvely. I(, j ) and I (, j) located at coordnate (, j ) of the orgnal mage and embedded mage. After extractng, estmaton for qualty of extracted watermark s compared to the orgnal watermark s express accordng the normalzed correlaton NC value: m n NC w(, j) w(, j) m n (4) j where s the watermark of sze. w(, j ) and w (, j) are the values located at coordnate (, j ) of orgnal watermark and extracted watermark. Value of w(, j ) s set f t s a watermark bt, otherwse, t s set - and smlar to w (, j). So the value of w(, j) w(, j) s ether or -. If number of bt extracted correctly s larger than 5%, the NC value s postve; otherwse, t s negatve. A. Wth Non-Attack In below the smulaton, the gray-mages used as cover and watermark s the mages have sze (5 5) bnary mage wth sze (6 3). Addtons, we apply the Haar wavelets to decompose and reconstruct to cover mage. Smulaton result of embeddng process showed n TABLE.II, all cover mages n ths smulaton get good nvsblty (PSNR larger than 4dB). In the non-attack case, the watermark mage s extracted from the cover mage accordng usng proposed determnng threshold algorthm. Table. II shows the qualty of recovered watermark wth NC= for all mages. TABLE II. RESULT OF EMBEDDED IMAGE AND RECOVERY WATERMARK Lena Baboon Pepper Goldhll Salboat Orgnal PSNR MSE..... B. Wth Geometrc Attacks Under varous attacks, we only use mage Lena for smulaton. There are some geometrc attack types are consdered n ths smulaton, such as, croppng, scalng, rotaton and Gaussan nose. For the croppng attack, pxel of mage s dropped at the center of embedded mage and they are replaced by gray value pxels. In the scalng case, the host mage s scaled to and then scale one more tme to orgnal sze. For two last attack types, the cover mage s rotated by small degree and added Gaussan nose wth many dfferent varances. The result of watermarkng under geometrc attacks s showed n Table. III. Wth the larger value T, the robustness of watermark wll be ncreased but the nvsblty of host mage wll be reduced. There s a tradeoff between the robustness of watermark and the transparency of host mage. TABLE III. Attack Type Croppng PSNR.67 NC NC OF EXTRACTED WATERMARK UNDER GEOMETRIC ATTACKS.64 Scalng C. Wth Non-Geometrc Attacks TABLE IV. Attack Type Hstogram Equalzaton PSNR 9. NC.88 Rotaton.5 Gaussan Nose NC OF EXTRACTED WATERMARK UNDER NON- GEOMETRIC ATTACKS Gaussan Flter Medan Flter (3 3) Average Flter (3 3) Sharpenng There are some non-geometrc attacks: Hstogram Equalzaton, Gaussan flter, Medan flter, Average flter, sharpenng. Based on the experment result represented n Table. IV, the proposed method has a good resstance to many non-geometrc attacks.

5 Internatonal Journal of Electroncs and Electrcal Engneerng Vol., No., June 3 D. Wth JPEG Compresson TABLE.V showed the smulaton result n JPEG lossy compresson wth qualty factor from to 6. When we compress the mage wth qualty factor %, although the sze on dsk of mage s very small, NC value of watermark s always greater than.57 and NC= for 6%. E. Comparng to The Other Methods TABLE V. Qualty Factor % PSNR NC NC OF EXTRACTED WATERMARK UNDER JPEG COMPRESSION % 3% 4% 5% 6% We compare our proposed method wth some recently publshed papers: Ln et al [3], Wang et al [5], Run et al [6]. In our method, the expermental result s far better than the lsted method, especally for JPEG lossy compresson, scalng, rotaton, low-pass flter, hstogram equalzaton. Only croppng and sharpenng, the proposed method s not good. Comparson s represented n TABLE.VI. TABLE VI. Attack Type Croppng ¼ Scalng Rotaton.5 Rotaton.3 Gaussan flter Hstogram Equalzaton Sharpenng Medan flter (3 3) Medan flter (5 5) Average flter (3 3) Average flter (5 5) JPEG (QF=) JPEG (QF=) JPEG (QF=3) JPEG (QF=4) JPEG (QF=5) COMPARING PROPOSED METHOD TO THE OTHER METHODS Ln et al [3] (PSNR=44.5dB) IV. Wang et al [5] (PSNR=45.9dB) CONCLUSION Run et al [6] (PSNR=4.98dB) Proposed (PSNR=43.dB) In ths research, we propose an effcent blnd watermarkng for dgtal mage when usng the adaptve threshold algorthm n extracton process. Ths threshold s gven based on computng weghted -wthn class varance through consderng probablty densty functon of sgnfcant dfference. The proposed method s consdered n smulaton under dfferent attack types. The proposed method n ths paper has good mperceptblty through PSNR peek sgnal-to-nose rato parameter (PSNR>4dB) and effectvely resst many attacks, such as, Gaussan flter, Medan flter, Average flter (NC - Normalzed Correlaton of extracted watermark larger than.95) and JPEG compresson (NC=.87 for QF=3%). However, ths method s weakly at some geometrc attacks: croppng, Gaussan nose. REFERENCES [] M. Hsel, D. Tseng, and Y. Huang, Hdng dgtal watermarks usng multresoluton wavelet transform, IEEE Trans. Ind. Electron, vol. 48, pp , Oct.. [] B. K. Len and W. Ln, A watermarkng method based on maxmum dstance wavelet tree quantzaton, n Proc. 9th Conf. Computer Vson, Graphcs and Processng, pp , 6. [3] W. H. Ln, S. J. Horng, T. W. Kao, P. Fan, C. L. Lee, and Y. Pan, An effcent watermarkng method based on sgnfcant dfference of wavelet coeffcent quantzaton, IEEE Transactons on multmeda, pp , Aug. 8. [4] W. H. Ln, Y. Rau, and S. J. Horng, A wavelet-tree-based watermarkng method usng dstance vector of bnary cluster, Expert Systems wth Applcatons, pp , 9. [5] Y. R. Wang, W. H. Ln, and L. Yang, An ntellgent watermarkng method based on partcle swarm optmzaton, Expert Systems wth Applcatons 38, pp ,. [6] R. S. Run, S. J. Horng, W. H. Ln, T. W. Kao, P. Fan, and M. K. Khan, An effcent wavelet-tree-based watermarkng method, Expert Systems wth Applcatons, pp ,. [7] M. Keyvanpour and F. M. Bayat, Bnd mage watermarkng method based on chaotc key and dynamc coeffcent quantzaton n the dwt doman, Mathematcal and Computer Modellng (), do:.6/j.mcm [8] A. Dng and S. Dong, Algorthm of Dgtal Based on Decomposton of Wavelet Packet, Sxth Internatonal Confer-ence on Internet Computng for Scence and Engneerng, pp ,. [9] K. Byun, S. Lee, and H. Km, A watermarkng method usng quant-zaton and statstcal characterstcs of wavelet transform, PDCAT 5. Sxth Internatonal Conference on Parallel and Dstrbuted Com-putng, Applcatons and Technologes, pp , Dec. 5. [] L. Fan and T. Gao, A Novel Blnd Robust ng SchemeBased on Statstc Characterstc of Wavelet Doman Coeffcents, n Proc. Internatonal Conference on Sgnal Processng Systems, pp. 5, May 9. [] M. Hajzadeh, M. S. Helfroush, and M. J. Dehghan, A Novel Blnd ng Method Based on Dstance Vector of Sgnfcant Wavelet Coeffcents, 6th Iranan Machne Vson and Process-ng (MVIP), pp. 6, Oct.. [] P. Lu and Z. Dng, A Blnd ng Scheme Based on Wavelet Tree Quantzaton, Second Internatonal Symposum on Electronc Commerce and Securty, vol., pp. 8, May 9. [3] J. Huang and C. Yang, Dgtal ng Algorthm Usng multresoluton Wavelet Transform, 4 IEEE Internatonal Conference on Systems, Man and Cybernetcs, vol. 3, pp , Oct. 4. [4] J. H. Uhl, C. Kodl, and A. Uhl, Multple Blnd Re-watermarkng wth Quantzaton-Based Embeddng, n Proc. 8th IEEE Internatonal Conference on Processng, pp ,. [5] Y. Qao, Q. Hu, G. Qan, S. Luo, and W. L.Nownsk, Thresholdng based on varance and ntensty contrast, The journal of the pattern recognton socety, vol. Pattern Recognton 4, pp , 7. Thuong Le Ten was born and grew up n Sagon, now s Ho Ch Mnh Cty Vetnam. He got Bachelor Degree n Engneerng from the Unversty of Technology, Vetnam n -98, snce then he has been wth the Electroncs & Telecommuncatons Department, Ho Ch Mnh Cty Unversty of Technology. He spent 3 years n the Federal Republc of Germany as a vstng scholar at the Ruhr Unversty of Bochum from 989 untl 99. He receved the Ph.D. n Telecommuncatons from the Unversty of Tasmana, Australa n 998. He served as Deputy Department Head for many years and has been the Telecommuncatons Department Head snce 998. He has also apponted for the second poston as the Drector of Center for Overseas Studes snce 998. Hs areas of specalzaton ncludes: Communcaton Systems, Sgnal Processng and Electronc Crcuts.

6 Internatonal Journal of Electroncs and Electrcal Engneerng Vol., No., June 3 Then Huynh The was born n Bentre provnce, Vetnam n 988. He receved the B.E degrees from the Unversty of Techncal Educaton Ho Ch Mnh Cty (HCMUTE), HCM Cty, Vetnam n. He s currently pursung the M.E degree at Faculty of Electrcal & Electroncs Engneerng, HMCUTE. Hs research nterests are sgnal processng, dgtal watermarkng, and hdng nformaton. 3